Long-running http backend panics 'native handle registration exhausted: IdExhausted' after ~200k requests #11453

Found by the tokio-removal measurement (#11451). It reproduces on both BEFORE (d8f24f15ed) and AFTER (bd8936e98e).

A long-running http backend (http server + fetch + crypto + zlib + timers) panics with common native handle registration exhausted: IdExhausted after about 200k requests in one process. Handle ids are being consumed and never returned. Any long-running Perry server will crash.

Also relevant: #11225 (merged) made dropped stdlib handle ids never reused and moved them into perry-ffi's shared pool (~262k ids). That is correct for avoiding stale-id aliasing, but it means ids are now strictly finite per process. Check:

1. which handles leak (never dropped) versus are dropped but never reclaimable;
2. whether the pool needs generation-tagged ids (reuse slots safely with a generation counter so stale ids don't alias) instead of never-reuse.

Repro: the backend probe in benchmarks/tokio-removal/ (#11451) under sustained load.