Fix og:title output; Handle list or string value

Browserify ALL THE THINGS

Use key derivation to provide secret keys

Add worker runner (hypothesis-worker)

Travis is being especially slow at the moment.

This brings bugfixes for infinite loop on FF.

Fixes 1983.

Fix notifications

Use karma-browserify to run the tests, rather than relying on prebuilt
assets. This commit alters our test harness so that it no longer relies
(at all) on the webassets pipeline. This means:

- test isolation is improved, as each test file runs in its own module
- bringing objects into test scope is now simply a matter of requiring
  the correct module, rather than bolting more and more objects onto the
  Annotator namespace
- autoreloading of tests when either test files or their dependencies
  are modified is now possible (and enabled by default)

This includes files that need to be shimmed under the "browserify-shim"
key. Annotator plugins that have no dependencies do not need to be
shimmed at the moment as they just attach themselves to the global
Annotator object.

jQuery is assigned as a global, this means that it is assumed to be
included separately and bound to the global scope.

Both jQuery and Annotator have aliases under the "browsers" key to
make them easier to reference.

Use a singular time expression one year

Close #1973

Removing some obsolete conditional code,
which was necessary for some obsolete version
of PDF.js.

Fortunately, by now, all supported versions
of PDF.js uses the same data format, so this
workaround can go.

Earlier, on some versions of PDF.js, we used to
use the PDF.js's FindController as a data source
for the text extraction.

However, at some point, we stopped using the
routines shipped with it, since it didn't always
provide use adequate spacing between the various
pieces of texts. So we ended up just contenating
the various pieces of text ourselves.

Then, for new versions of PDF.js, we introduced
other means of accessing the same information,
completely bypassing the PDFFindController.

This change simply unifies the access; now we can
do the same an all PDF.js versions.

Allow using the API as a virtual root

Add highlight test and change scrolling

... by moving some repeating code into
forEach blocks of the relevant sections.

Fix real-time delete

When receiving a real-time notification about the removal
of an annotation, we emit an annotationDeleted event,
so that all components can react.

Earlier, we emited this event simply with the Annotation
bject that arrived on the wire.

However, some components couldn't deal with this, because
they were expecting to see the 'real' Annotation object,
the ones they already knew about.

So now we do a lookup before emiting the event, and use
the locally found objects instead.

Standardise on userid

This resource request adds a `user_id` query param. But the
corresponding view function doesn't pay any attention to such a
parameter, simply inspecting the authenticated userid to work out whose
profile to provide:

  https://github.com/hypothesis/h/blob/15c39c8b58a8800d7a985c1d29b83fbdb6873266/h/accounts/views.py#L254

Disable extremely slow auto-update of ancient timestamps

When a timestamp is several months old, and we leave
the application around (in a forgotten tab) for
more than three weeks, the usual method of updating
the timestamp fails.

This change works around that by maxing out our timeouts
at 24 days. (This will simply cause us to check the
time in every 24 days.)

Fixes #1952.

I also moved the 5-second minimum to the same time helper.

The goal of this change is to avoid any false appearance
that we are implementing the proposed scrollIntoView() API [1],
while we clearly don't.

So, changes:

 - Rename scrollIntoView to scrollToView
   on both Anchor and Highlight
 - Stop returning promises at both methods

I also updated all the tests to reflect this change.

[1] http://dev.w3.org/csswg/cssom-view/#dom-element-scrollintoview

Load all annotations

To avoid temporary "message not available" cards

Instead of just one search call (with a fix limit),
process the total field from the result of the first
load and while there is still more to load, initiate
a new search query, with the offset property set to
load the remaining ones