serialize: reduce header-path allocations safely

[anmonteiro]

Summary

• avoid List.rev allocation on hot header paths by iterating headers in HPACK/wire order (Headers.iter_hpack)
• use a reusable header scratch encoder in Serialize.Writer instead of allocating a fresh Faraday for every request/response/trailer header block
• preserve correctness for queued writes by copying header iovecs into the main writer for HEADERS/PUSH_PROMISE/CONTINUATION payloads (avoids shared-buffer aliasing)
• add a TODO documenting follow-up work to move to pooled header buffers for zero-copy safety

Why

This targets two perf findings:

• repeated header-list reversals in serialization/validation
• per-message Faraday allocations in header encoding

The initial shared-scratch zero-copy attempt regressed an existing test (headers blocks don't share buffers) due buffer aliasing. This PR keeps the optimization while maintaining correctness.

Validation

• dune runtest --no-buffer