Split FreeWithHooksOrPerThread into per-thread and hooked free slow paths.

[copybara-service]

Split FreeWithHooksOrPerThread into per-thread and hooked free slow paths.

FreeSmallSlow routed both "delete hooks installed" and "per-CPU caches
inactive" into one NOINLINE callee. That callee unconditionally built a
DeleteInfo (class_to_size load + 4 stack stores), tested delete_hooks_, then
re-tested CpuCacheActive before falling into the inlined ThreadCache::
Deallocate. Per-thread-mode binaries therefore paid the hook plumbing on
every free, and the hooked per-CPU path carried the ThreadCache body in its
frame (3 callee-saved pushes, 0x38 stack).

Dispatch directly from FreeSmallSlow instead:

• HaveHooks() -> FreeSmallHooked (hook + DeallocateSlow, or the
  per-thread callee if per-CPU is inactive),
• !UsePerCpuCache() -> FreeSmallPerThread (ThreadCache::Deallocate /
  transfer cache only, no hook plumbing),
• otherwise -> inlined DeallocateSlowNoHooks (unchanged).

Both are tail calls out of FreeSmallSlow; its hot per-CPU body is unchanged
(register renaming only). Routing still keys on HaveHooks() rather than
delete_hooks_.empty() because new-hook-only installs still need
DeallocateSlow -> MaybeForceSlowPath to uncache the slab. Sampled-object
handling is untouched. The FreeSmall inline body and fast_path goldens are
byte-identical.

x86-64 -c opt, per-thread free (thread cache present, no overflow), callee
after FreeSmallSlow's dispatch:
before FreeWithHooksOrPerThread: 51 insns, 15 loads, 10 stores, 4 pushes
after FreeSmallPerThread: 32 insns, 9 loads, 5 stores, 1 push
Hooked per-CPU path: same instruction sequence, one fewer ptr spill/reload;
function shrinks 402 -> 249 bytes. FreeSmallSlow +16 bytes (second tail
call).