Remove the ban on using console.log() in tests

Simplify request preparation in `store.js`

This was intended to prevent console.{log,warn,...} calls accidentally
being left in test code.

However the result was very confusing for H developers who only
occassionally worked with the H client test suite.

This problem is better solved by a linter.

Upgrade PhantomJS from 1.x to 2.x

Fix new replies being moved to the newly focused group

Use "log in"/"log out"/"sign up" consistently

Unless there is a clear reason not to do so, it is generally preferred
to use methods from the ECMA standard lib.

Rather than using a $http request transform to preprocess the data,
just pass the data object into the `stripInternalProperties` function
directly.

Previously the `group` and `permissions` properties of new replies were
set _after_ creating the reply with `createAnnotation()`. This meant
that the new reply did not have the correct group and permissions when
it was filtered for the first time by the `buildThread()` function.

Work around a bug in Karma 1.10 [1] which caused console messages from
tests not to be propagated to the terminal.

A fix has been accepted upstream and this workaround can be removed
when the next release of Karma is made.

[1] See https://github.com/litixsoft/karma-mocha-reporter/issues/67

Under PhantomJS 2.x `new URL(uri)` works, so the feature test in
`polyfills.js` passes, but the resulting URL object is an apparently
empty object.

Properties such as `hostname` and `href` are missing for example.

Load the URL polyfill explicitly in the Karma-specific polyfill module.

Under PhantomJS 2.x, the output of `Date.toLocaleTimeString()` includes
'GMT+01:00', where the '+' needs to be escaped.

 * Upgrade Karma from v0.13 -> 1.10

 * Upgrade from PhantomJS 1.9.x to latest 2.x

This fixes an issue where the gulp process would hang after tests ran
when using a current version of Node.

Only new top level annotations should be moved to the new group when
switching groups.

Fixes hypothesis/h#3559

Simplify annotation component by removing duplicated state

Remove site scripts and assets

The correct fix would be to remove new annotations with an empty draft
in the reducer function that handles new annotations being created. That
is not currently possible because drafts are not stored in the Redux
store. Therefore this commit just manually de-registers the listener
when <annotation> goes out of scope.

The `updateViewModel()` function was responsible for computing several
pieces of data (links, document domain and title) used by the annotation
template.

Since we can now rely on vm.annotation being immutable, we can just
compute this data lazily when we need it and use memoization to avoid
unnecessary recalculation.

Replace `updateDomainModel()` with a simpler function which returns a
copy of the input annotation with the changes from the editor applied.

These variables duplicated the information from the current draft for an
annotation. Instead of restoring a draft when an annotation card is
created and saving it when the annotation card is destroyed, just update
the draft whenever the tags, text or privacy of the annotation is
changed. The text/tags/privacy shown in the card are determined by
`vm.state()` which merges the annotation and the draft.

Removing these variables lets us also remove or simplify the tests that
checked that everything remained in sync.

Instead of maintaining a state flag indicating whether an annotation is
being viewed, edited or created, derive that state from whether the
annotation has an ID and whether it has an unsaved draft or not.

In the process this commit simplifies the tests for reverting edits and
adds a missing test that the annotation is deleted if new when clicking
the Cancel button.

As a step towards making <annotation> stateless or nearly-stateless and
instead just a display of an annotation derived from the current
annotation instance and its draft, remove the `domainModel` copy of
`vm.annotation`.

Make `gulp watch` build CSS, files and images

Make `gulp watch` build the CSS, files and images initially (creating
the output files if they don't exist yet, overwriting them if they do)
and then watch the source files and re-build them when they change.

This matches the existing `gulp watch` behaviour for scripts.

This means you can just run `gulp watch` and have everything built,
instead of having to run `gulp build watch`.

Fix annotation state not being updated when annotations are saved

Annotations in the app's local state store were not being updated
after a save was successfully committed on the server.

Fixes hypothesis/h#2965

zen-observable update

Remove use of ngResource for search and annotation queries

Show direct links on stream page when available (FIRST!)

The CHANGES file is part of the 'h' repo and no longer exists in the
'client' repo that was forked from it.

Move relevant tests to annotation-metadata-test.js, and fix a bug
in annotation-metadata.js.

Several aspects of ngResource make it sub-optimal for our needs:

 1. It mutates the model object directly after making an API call, which
    does not fit well with usage in a Redux app where the UI state should
    be an immutable object.

 2. The ngResource classes can only be constructed once the API
    description has been retrieved. At least one place in our code,
    which handled newly created annotations arriving from the page,
    failed to account for this.

This commit therefore replaces use of ngResource for making API calls
to the search and annotation endpoints with a simple wrapper around HTTP
requests which makes a call and then returns the resulting object.

The new wrapper will wait until the API description has been received
before attempting to make an API call, avoiding problem (2) above.

As a bonus, this brings us a step towards decoupling the JS API client
from Angular so we could re-use it outside the app in future.

 * Replace ngResource with simple $http wrapper in store.js

 * Add tests for annotation update and deletion API calls

 * Change tests in annotation-test.js from
   `assert(predicate(actual, expected))` form to
   `assert.predicate(actual, expected))` form as this results in errors
   that are _much_ easier to debug when they fail.