Previously, this would issue a query for each unique `diff_refs_or_sha`
passed. This was because we didn't want to load other MR diffs into memory, as
they had some very large columns.

Now they are actually very small, and it's more efficient to just load them all
at once and do the finding in Ruby.