Use a ForkJoinPool instead of a PriorityBlockingQueue-based ThreadPoolExecutor
for the analysis phase. During the analysis phase, lots of very short actions
are executed (File, FileState, ...) and contention in the executor has a
substantial performance impact.

Contention has always been a problem here, but the move towards rightfully
using PriorityBlockingQueue can cause even more work to be executed under a
lock. While I couldn't measure substantial wall-time differences between
PriorityBlockingQueue and the previously used LinkedBlockingQueue, this change
improves performance by about 10% on a variety of experiments I have
conducted.

RELNOTES: None.
PiperOrigin-RevId: 210398382