Hide "Public" group for unauthenticated users (alternate implementation)

Implement a business requirement that the "Public" group should be
hidden from the groups list if all these conditions are met:

 - The user is logged-out
 - The current document has groups associated with it
 - The user has not visited a direct-link to an annotation in "Public"

---

This requirement introduces a complexity in that deciding which group to
focus and fetch annotations for after initially fetching groups now
depends on the current profile. However the profile and groups are
fetched concurrently to reduce the number of API round-trips before we
can show annotations in the client.

Fortunately we don't need to know the exact userid to know which group
to focus, only whether the list of groups was fetched as a logged-in
user or not. This can be determined from the access token that was sent
with the `/api/groups` request.

Remove obsolete JSDoc comments

Add an option to return metadata in the response as well as just the
JSON response from the server. This will initially be used to determine
the access token that was used to make the request and therefore whether
it was authenticated or not. In future it could also return information
from HTTP headers or the status code if useful.

Remove unnecessary extra `/api/groups` call on startup

Respond to PR feedback.

This event is never broadcast anywhere. A comment referenced this being
called when a user leaves a group, but this is handled in the `streamer`
service by calling `groups.load()` directly.

On startup an initial `FRAME_CONNECTED` event is fired when the main
frame connects to the sidebar. This triggered an extra `/api/groups`
fetch. Avoid this by recording the main frame URI used in the last
groups fetch in  `load()`. The existing test in the `FRAME_CONNECTED`
handler then skips the `load()` call if this URI has not changed.

The group service does not have `add` or `remove` methods any more.

Fix `groups.focused()` error if localStorage is empty when client loads

Use a pointer cursor for highlights

If the client started and local storage did not contain a saved
last-viewed group ID, then `groups.load()` would attempt to call
`store.focusGroup(null)`. As a result subsequent calls to
`groups.focused()` would return null and code in various places in the
application expects that there will always be a focused group.

Fix the problem by not setting the focus group after groups are fetched
unless the previously-focused group was set and exists in the new
groups list.

Also improve the behavior of `store.focusGroup` by making it leave the
focused group untouched if the specified group has not been loaded, but
log an error so we'll know (via Sentry) if it happens again.

Fixes #750

The `annotator-hl` class contains a `cursor: pointer;` rule already, but
it only applies to mobile devices (more precisely, devices which lack a
sufficiently accurate primary pointer, like a mouse). I simply removed
this restriction so that the rule applies across the board.

Closes: hypothesis/product-backlog#703

Convert share dialog links to use bouncer instead of via.

If the URL is not encoded, a frame URL such as
"http://example.com/foo&bar=baz" will generate the page share link
"https://hyp.is/go?url=http://example.com/foo&bar=baz" where the
"&bar..." part becomes a separate parameter instead of part of the "url" param.

Also remove the check to see if the user is already on a via page.
Via inserts a <link rel="canonical" href="{original URL}"> tag in the
proxied page and therefore frames[0].uri is the original URL and never
https://via.hypothes.is/{original URL}

Fixes https://github.com/hypothesis/product-backlog/issues/335

Fix warnings about `Buffer` constructor during `gulp build`

Fix warning about deprecated `Buffer` usage in Node 10 at the end of
`gulp build`.

See https://nodejs.org/api/buffer.html#buffer_buffer_from_buffer_alloc_and_buffer_allocunsafe

`externalContainerSelector` config implementation

Fix various failures when running test suite under Chrome

Fix 404 warnings from headless Chrome during tests caused by trying to
load non-existent image URLs in the anchoring regression tests.

The previous method of setting the `__hypothesis_frame` property on the
Window did not work. I haven't investigated in detail why, but the
alternate method of setting this property used here works instead.

Fix adder toolbar tests in browsers (eg. Chrome) which natively support
Shadow DOM v1. The test for Shadow DOM v0 failed because the implementation
feature-detected and used the Shadow DOM v1 method
`Element.prototype.attachShadow` instead.

Fix the issue by patching the test element to hide the native Shadow DOM
methods if they exist.

`uri` is a mandatory property of actual annotations. The absence was
causing some tests to fail in Chrome due to the URL constructor being
called with `undefined`.

`toLocaleTimeString()` in Chrome produces an `hh:mm:ss` output on my
system whereas the rendered timestamp only has `hh:mm`.

Re-using `format` from date-util avoids this hazard.

Add missing arguments to `Element#setAttribute` and
`Selection#collapse`.

PhantomJS allowed these missing arguments but other browsers, eg.
Chrome, do not.