Count graph nodes per address bucket so ForgetDeadlockInfo() can skip deadlock_graph_mu.

In debug builds with deadlock detection enabled, ~Mutex() calls
ForgetDeadlockInfo(), which takes the global deadlock_graph_mu spinlock to
remove the Mutex's node from deadlock_graph. Only Mutexes involved in nested
locking get a node, so for most destroyed Mutexes there is nothing to remove.

Keep a count of deadlock_graph nodes for each hash bucket of Mutex addresses.
ForgetDeadlockInfo() reads the count for the Mutex's bucket without the lock,
and returns early if it is zero:

- Counts change only under deadlock_graph_mu, when GetGraphIdLocked() creates a
  node and when ForgetDeadlockInfo() removes one. GraphCycles::GetId() can now
  report whether it created a node, and RemoveNode() returns whether the node
  existed.
- If the Mutex has a node, the DeadlockCheck() call that created it either
  happens-before ForgetDeadlockInfo() or races with it. In the first case,
  ForgetDeadlockInfo() sees a nonzero count. In the second case,
  DeadlockCheck() could also create the node just after the removal, so the
  early return adds no new race.
- Counts saturate at 255 and then never change, so a saturated bucket is never
  treated as empty again. Counts below 255 are exact.

There are 2^18 buckets, with 8-bit counts packed four per 32-bit atomic word.
8-bit counts give 4x as many buckets as 32-bit counts in the same memory, so a
larger fraction of buckets is empty. Packing keeps the array
constant-initialized: an array of 2^18 one-byte atomics exceeds the compiler's
constexpr evaluation step limit for ABSL_CONST_INIT.

Cost: a 256 KiB zero-initialized static array, touched only in debug builds
with deadlock detection enabled.
PiperOrigin-RevId: 997581888