Background information

Currently, the bucket-bar should display anchor positions only from a
_single_ guest frame. This is because there is no merging mechanism for
anchor positions from multiple guest frames.

Before this PR

* The host frame listened for `anchorsChanged` events from
_all_ the guest frames.

* Only _one_ guest frame sent this event. We referred to this guest
  frame as having the 'main' annotatable content.  The 'main'
  annotatable guest frame was identified by not having a
  `subFrameIdentifier` (a configuration option added to a guest frame
  when the Hypothesis client was injected). Hence, the 'main'
  annotatable guest frame was always the frame where the Hypothesis
  loaded initially (in contrast with the injection mechanism).

In this PR

We have reversed the logic of how the `anchorChanged` RPC
events are send and received:

* _Every_ guest frame sends `anchorsChanged` RPC events to the host
  frame.

* The host frame chooses to listen to only _one_ guest frame for the
  `anchorsChanged` RPC events: the guest frame that connects first with
  the host.

For non-VitalSource case, the first guest frame that connects to the
host frame is the one where the Hypothesis client was initially loaded
(doesn't contain `subFrameIdentifier` option).

For the VitalSource, because of #4176, there is no guest frame in the
host frame. This allows the guest frame in the book content, where the
Hypothesis client was injected, to send anchor positions, and positions
to be reflected in the host's bucket-bar.