Implement per-size-class lock sharding to reduce contention

[bpowers]

• Replace single _miniheapLock with per-size-class _miniheapLocks array
• Add separate _largeAllocLock for large allocations (sizeClass == -1)
• Add _arenaLock to protect shared arena/allocator state
• Add AllLocksGuard RAII class for acquiring all locks in consistent order
• Make _miniheapCount atomic for safe access under different locks
• Update freeFor() to use per-size-class locks with proper ordering
• Defer flushBinLocked() to meshing cycle to avoid lock ordering issues
• Update mallctl(), getSize(), okToProceed() for new lock structure

Lock ordering: arena → large → sizeClass[0..N-1]

This idea is that this can reduce lock contention in multi-threaded workloads by allowing concurrent operations on different size classes.

[emeryberger]

@bpowers is this still in progress?