Speed up and slightly simplify Mutex

Eliminate the separate seqentially consistent contention counter load
in Mutex::ExclusiveUnlock by putting the contenion counter in the
Mutex's state word.

Replace some CHECK_GE checks with CHECK_GT. We were checking quantities
intended to be non-negative against >= 0 just before decrementing them.

Remove a pointless volatile declaration.

Introduce constants for the first FUTEX_WAKE argument. Remove all uses
of -1 as that argument, everywhere in ART. It appears to work, but the
documentation says it's wrong.

This does not yet address the ReaderWriterMutex issue, which is handled
in a different way in a separate CL.

Benchmark runs with and without this CL weakly suggest a tiny, not
statistically significant, improvement in both time and space with this

Bug: 111835365
Test: Build and boot AOSP. TreeHugger.
Change-Id: Ie53c65f2ce774a8cb4d224e2c1b3a110eb880f0c
4 files changed