[clangd] Run code completion on each token coverd by --check-lines

In --check mode we do not run code completion because it is too slow,
especially on larger files. With the introducation of --check-lines we
can narrow down the scope and thus we can afford to do code completion.

We vlog() the top completion result, but that's not really the point.
The most value will come from being able to reproduce crashes that occur
during code completion and require preamble build or index (and thus are
more difficult to reproduce with -code-complete-at).

Differential Revision: https://reviews.llvm.org/D103538