pool/multislab: erase is O(S) #5

pool::erase (pool.cppm:389) already holds the node in its iterator but calls multislab::deallocate(void*), whose find_owner (multislab.cppm:583) scans every slab. Slabs are fixed size, so S grows linearly with size. emplace's rollback and make_iterator(ptr) pay the same scan.

Depends on: multislab list rework (stable node handle in the iterator).

Fix

1. Node-aware deallocation: private multislab::deallocate_at(const iterator&) with pool as friend, or public deallocate(iterator). Node and bit index are known, so there is no scan and no ptr_to_index division. Use it in pool::erase and in emplace rollback (allocate_at already returns the iterator).
2. Optional, separate issue: O(1) deallocate(void*) by allocating each slab at power-of-two alignment with the header at a fixed offset, so the owner is ptr & ~(slab_bytes - 1). Needs aligned_memory_resource and a padded footprint.

Update docs/complexity.md and the find_owner note in docs/containers.md.

Tests

Erase benchmark over 10k+ slabs. Unit test: erase through deallocate(iterator) leaves the same state as deallocate(ptr).