We limit the number of candidates returned to Vim to 20 and also make sure that
we are not returning any duplicate candidates. This provides a noticeable
improvement in latency.
First off, we don't block the GUI thread anymore for ClangCompleter (that was
always temporary). Secondly, now ClangCompleter will cache the data coming from
clang so that query-based filtering of members is fast.