perf: skip handle registration for non-escaping createHash/createHmac chains (removes #11515's GC-trace cost) #11516

Follow-up to #11515 (fix for #11453, handle-id exhaustion).

#11515 makes crypto hash/HMAC/cipher/sign/StringDecoder handles GC-reclaimable, proving an id unreachable with full traces before reuse. That fixes the crash after ~200k requests (1.6M requests now served with flat RSS) and cuts backend RSS by 43%. The cost is extra full-trace work:

• +2.6% instructions on the http backend probe;
• +9.8% on a loop that only does createHash.

The base build avoided that cost only by leaking every handle until it crashed.

To remove the remaining cost without giving back the memory win: have the compiler recognise non-escaping handle chains such as createHash(alg).update(x).digest(enc) and createHmac(...).update(...).digest(...) (the dominant per-request pattern). Then either:

• skip registering a handle at all (compute the digest directly from a stack-held state); or
• release the id immediately at the chain's end, because escape analysis proves no JS reference survives.

Measure with the #11515 microbenches under /tmp/perry-bench-lock.d; the target is the pre-#11515 instruction count with #11515's bounded RSS.