Cache one freed block per size class in pymalloc #159208

I'd like to try keeping one freed block per size class when it comes from a full pymalloc pool. The next allocation of that size can reuse it without updating the pool's count or adding and removing the pool from a list. Full GC and interpreter cleanup return cached blocks to their pools. Public APIs and allocator hooks stay the same.

Linked PRs

• gh-159208: Cache one freed block per size class in pymalloc #159209