Fixes #2371
Adds test.

Search for one URI

Introduce the h.api.uri module for URI normalisation/expansion

Previously, when loading Hypothesis into a page, the client would make
one search request for each and every URI found on the page, including
those found by the document metadata scanners (in the Annotator
"Document" and "PDF" plugins.)

This is wasteful of network bandwidth, and moreover should not be
necessary in most cases due to the ability to expand to equivalent URIs
on the server side.

The component responsible for coordinating communication across page
frames is the "crossframe" service. This services maintains a registry
of page frames (previously called "providers") and their associated
metadata. This commit reduces the metadata stored to simply the page
URL, or in the case of PDF documents, the fingerprint URN.

It's worth noting that this *does* change behaviour in one important
way: if a page contains document equivalence data that we've never seen
before, and one of the alternate URIs has already been annotated, then
this change will mean we do not load annotations until that page itself
has been annotated (and thus the document equivalence data updated). I
am assuming that this is an extreme edge case.

Anchoring rewrite

s/viewable/visible/

Prevent annotation attempts when not signed in

Simply hides the annotation form--which implied you could
use the thing...--and instead tells the user to sign in,
so they can do what they came to do.

While transforming a selector to a range must occur synchronously,
the individual strategies can be attempted in separate animation
frames.

Also, make this more beautiful.

All of the anchoring types are actually synchronous, for now. Having
promise callbacks between the selector->anchor and anchor->range
conversions sometimes creates race conditions when multiple
annotations anchor concurrently. Punt on this for now.

And hide the cache details from the default implementation that
doesn't use them.

I'm not sure this works because I'm not sure the first then function
will be called synchronously by the render frame invoking resolve.

This makes it more robust against change by storing its description
at the container level rather than a DOM Range object that could get
mutated unpredictably by text node splitting.

Rather than searching PDF pages in order, use the position anchor
(when available) to spiral the search outward from the page that
likely contains the quote.

Rather than searching all the pages in parallel, search them serially
and terminate early with a match.

This is maybe too verbose for now but at least it's explicit and
doesn't assume that anchoring implementations can be fully
responsible for cache invalidation.