When indexing review requests, we determine if a review request is

accessible via review_request.is_accessible_by(). Part of this checks

the files in the diffsets for accessibility, using ACL diff checks. This

ends up querying the list of diffsets for each review request and then

each file within it, which drastically slows down diff indexing.
We now employ a couple layers of protection against this:


ReviewRequest.get_diffsets() will now use the prefetch_related()

   and select_related() caches to determine if it needs to even

   perform a new query, or if it can short-cut the result. This saves

   the per-diffset and prefetched-files queries.


ReviewRequest._are_diffs_accessible_by() no longer even calls

get_diffsets() if there aren't any FileDiffACLHooks registered,

   which will be the common case.


This ensures we don't do any more work than we need to do at any stage

of these checks, and makes a major difference in indexing performance.

All unit tests pass.
Added SQL-level debugging for the database backend and performed a

search index. With get_diffsets() being invoked for every diff, I

verified that it used the prefetch cache, avoiding new queries. With

ACL hooks factored in, I verified it never even got to the diffset

query code without faking hooks being available.