fix: TOCTOU race in and_compute_with under multi-threaded tokio

[Squadrick]

In try_compute and try_compute_if_nobody_else, the waiter was removed from the waiter map (via set_waiter_value(ReadyNone)) BEFORE the actual cache mutation (insert_with_hash/invalidate_with_hash). This created a window where a concurrent and_compute_with caller could:

1. Successfully insert its own waiter (the first was already removed)
2. Read the cache and see the OLD value (first caller hasn't written yet)
3. Both callers execute Op::Put based on the same stale data

This caused silent data loss under concurrent access with a multi-threaded tokio runtime (e.g. 10-15% of increments lost in a concurrent counter test with 8 writers).

The fix defers set_waiter_value(ReadyNone) to after the cache mutation in each match arm, ensuring the waiter remains in the map (blocking concurrent callers) until the cache is updated.

Contrast with try_init_or_read (used by get_with), which already does this correctly: it calls insert_with_hash BEFORE set_waiter_value.