Shrink ART Mutex exclusive_owner_ field to Atomic<pid_t>

The old volatile uint64_t version had a data race, and was thus
technically incorrect. Since it's unclear whether volatile uint64_t
updates are actually atomic on 32-bit platforms, even the informal
correctness argument here already effectively assumed that the upper
32 bits were zero. Don't store them. Explicitly complain if a pid_t
might be too big to support lock-free atomic operations.

Remove many explicit references to exclusive_owner to avoid
littering the code with LoadRelaxed calls.

The return convention for GetExclusiveOwnerTid() was unclear
for the shared ownership case. It was previously treated
inconsistently as 0 (pthread locks), (uint64_t)(-1U) and
(uint64_t)(-1). Make it as consistent as easily possible, and
document remaining weirdness.

Bug: 65171052

Test: AOSP builds. Host tests pass.

Change-Id: Ia99aca268952597a90b3c798b714cddbdc2c365e