ARC: shard the weak-reference table

[DTW-Thalion]

The weak-reference table took a single global lock, so weak operations on unrelated objects serialised on one mutex and scaled negatively with core count.

The striping now lives inside the table type rather than at its call sites: WeakRefTable holds an array of independently locked stripes, picks one from the object address, and keeps the operations it offered before, so nothing outside it knows the table is striped at all. The stripe count is a build flag or the OBJC_WEAK_SHARD_COUNT environment variable.

Three things the striping needed in its own right: the stripe a block belongs to is fixed when the block is created, which closes a lock-selection race that showed up as a use-after-free under AddressSanitizer with more than one stripe; a block's weak slot is accessed atomically, so the striping also holds on weakly ordered targets, checked under ThreadSanitizer; and a block is recycled within its own stripe.

#399 has merged, so the diff here is this change alone.

[davidchisnall]

This looks more invasive than it should be. The original refactoring to make the type used for a weak ref table distinct was intended to make it easy to provide a replacement that did the sharding internally.