These are used when the datetime in question is more than 24 hours ago
but still in the current year, and when it's not in the current year,
respectively. They replace a few different datestring formats that were
previously being used for datetimes in this range.

This produces "1 days ago" which is wrong, will be fixed later.

In date strings on annotation cards.

In date string formatting functions. Will enable more diverse
formatting.

In time.js replace format strings like '{} minutes ago' with functions
like nMinutesAgo(delta) that return strings.

The format strings only support datetime string formats of the type "n
seconds/minutes/hours/days/months/years ago". Although we don't take
advantage of it yet with format functions we can format datetime strings
any way we want, for example "16 April" for a date more than 24 hours
ago or "16 Apr 2016" for a date more than a year ago.

Although we're not taking advantage of it yet, this should allow us to
use more complex logic such as "within the current year" to select a
date format.

This is the first part of the new threading UI implementation
whose design is outlined at https://github.com/hypothesis/h/pull/3176/

build-thread implements the logic for constructing an immutable view
model for a list of conversations (a 'Thread') from a list of
annotations and replies plus an object describing the current UI state.

This view model is then rendered by an <annotation-thread> component.

Once the other parts of the new implementation are ready, this will
replace 'threading', 'thread-filter' and the JWZ script.

Make nipsa work with Postgres feature flag

Sync PG annotations to ES

Using custom element names for the adder's container is a simple trick
to reduce the likelihood of the adder's styling being affected by CSS
selectors in the containing page, such as example.com's `div` selector.

Remove compass-mixins dependency

Build Firefox WebExtensions during Travis builds

This is a global available in the extension and content scripts.

 * Use wildcard origin when sending messages to or from app.html
   served from a background page.

 * In embed.js, access the 'chrome' global via 'chrome' rather than
   window.chrome. In Firefox the 'chrome' global is not attached to
   the `window` object.

 * Work-around difference in how content script results are reported
   in Firefox (an object) vs Chrome (an array of objects, one per
   frame).

 * Avoid trying to call APIs which do not exist in Firefox.

Publish events in response callback

The immediate problem this dependency causes is that it it fails to
build with the current version of node-sass (see
https://github.com/Igosuki/compass-mixins/issues/84) which in turn means
that H cannot be built with Node 6.0. Removing this dependency allows
the node-sass dependency to be unpinned so that the current version
can be used.

The primary functionality that compass-mixins was used for was vendor
prefixing, which is now taken care of by autoprefixer, plus CSS fixes
for browsers which we no longer support (IE < 10).

For the remaining handful of mixins for resets and color mixing which we
are still using, this commit just extracts the few lines of code for
each from the compass documentation and includes it directly in our SCSS
files.

Convert AnnotationViewerController to JS

Show adder if user makes a selection and then launches the plugin.

Clean up reply notifications code

Make client test page load embed from same hostname as test page

See https://blog.readthedocs.com/securing-subdomains/

Try to avoid showing the adder above the selected text where possible by
showing the adder above or below the selection depending on whether the
selection was made forwards, in which case the focus point is assumed to
be at the bottom of the selection, or backwards, in which case the focus
point is assumed to be at the top of the selection.

If the last line of the selection is very close to the top of the screen
then show the adder pointing up at the text rather than down at it.

Also constrain the adder's position so that it never appears outside the
viewport.

Extract the adder styles and logic into separate files for easier
testing and extension.

Also extract the styling for positioning the arrow out into a a helper
class so we can more easily reposition the arrow depending on the
available space.

If a user selects some text, and then realises that Hypothesis isn't
active, their first reaction will probably be to activate Hypothesis.
In this scenario, we should pop up the adder near the selection on launch.

 * If there is a selection when the 'ready' event fires, show the adder
   at the focus point of the selection.

 * For consistency, show the adder at the focus point of the selection
   after a selection whilst Hypothesis is active.

 * Add 'range-util' module which provides utility functions to get
   the coordinates at the end of the selection.

   'range-util' uses the DOM Range, Selection and NodeIterator APIs directly
   rather than Annotator's Range functions for several reasons:

   1) Given that H now requires IE >= 10, all of our target browsers
      have all the APIs that we need.

   2) Using the DOM APIs directly proved easier to debug and test.

Fix front-end build - pin node-sass dependency to v3.4.x

Fix build by pinning node-sass dependency (used indirectly via
gulp-sass) to v3.4.x until
https://github.com/Igosuki/compass-mixins/pull/86 is resolved.

Previously the live reload server at http://localhost:3000 hardcoded the
embed script to 'http://localhost:5000' which would cause the embed to
fail to load if accessing the live reload server via a different IP.

This happens when accessing the server through a VirtualBox Guest when
using the default networking settings or when accessing it from a say, a
mobile device on the same WiFi network as a laptop serving the dev
environment.

Instead use the same hostname for the live reload WebSocket server
and embed script as the client test page is served from.

Update front-end testing dependencies to current versions and switch to mocha reporter

Fix a couple of slow front-end tests and enable slow test reporting

This should help to catch slow tests in development before they become
flakey failures on Travis.

 * Avoid `assert.includes()` in Bridge test, since this does a deep
   equality comparison which is much more expensive than necessary
   (50-60ms to execute on a desktop system with no contention. I think
   this might explain the very occassional failure seen on Travis).

 * Use the correct name for the `retries` option in a
   retryPromiseOperation() test. This test should have failed but
   because the min timeout value was set so low, the test was taking < 2
   seconds to run even though it was retrying 10 times with an
   expontentially increasing timeout.

When running tests in `gulp watch` mode, the initial run passed but
subsequent changes resulted in an error in the `merge-descriptors`
package, related to PhantomJS.

I haven't looked deeply into the problem but it does look like a
familiar issue related to Proxyquire trying to 'call through' to the
original imports in the module. Adding the '@noCallThru' flag in
guest-test.coffee resolves the issue.