Typing in replies in large threads (eg. those that already have hundres
of replies) was slow because every keystroke would cause every
reply to be re-rendered. The critical thread list virtualization
optimization in the client only applies to top-level threads and doesn't
benefit the case of a single thread with a large number of replies.

Implement some easy performance optimizations for large threads (ie.
with hundreds of replies) by memoizing:

 1) `Thread` components so that they don't re-render if something
    changes in a parent but the thread itself is unchanged.
 2) `Annotation` components so that they don't re-render if something
    changes elsewhere in the thread list but the current annotation is not
    changed.

As a result, creating and editing replies in large threads is faster. The
initial action to show a large thread list by clicking "Show replies" is still
slow because this change doesn't help the initial render. However,
editing performance is the most important issue from a user's
perspective.

Partial fix for #2897