perf: pool scratch header encoders and use zero-copy payload scheduling

[anmonteiro]

Summary

• replace copy-based header payload writes with schedule_bigstring-based iovec scheduling for HEADERS, CONTINUATION, and PUSH_PROMISE payloads
• add a small pool of scratch header encoders in Serialize.Writer
• only reclaim in-flight scratch encoders once writer output is fully drained, preventing buffer aliasing while avoiding per-message copies

Testing

• dune build --no-buffer @runtest-test_h2 @runtest-test_h2_client @runtest-test_h2_server @runtest-test_priority