perf: batch inbound WINDOW_UPDATE frames

[anmonteiro]

Summary

• batch inbound WINDOW_UPDATE credit instead of emitting one frame per body read callback
• keep connection-level credit pending across streams and batch per-stream credit within each active body
• update flow-control tests to cover both delayed small reads and threshold-crossing batched updates

Why

On large uploads, h2 was returning flow-control credit on every consumed chunk at both the connection and stream levels. Under the 1.5 GiB curl upload benchmark, that creates heavy write-loop churn and a large number of control frames.

Validation

• dune build --display=short @runtest-test_h2 @runtest-test_h2_client @runtest-test_h2_server spec/eio_h2spec.exe
• all tests passed

Benchmark

Using the existing 1.5 GiB curl upload benchmark against the Eio server, with the same local gluten checkout in both variants:

• baseline (cb71ce5): 4.168s, 368.52 MiB/s; 4.156s, 369.57 MiB/s
• batched WINDOW_UPDATEs: 3.231s, 475.41 MiB/s; 3.175s, 483.73 MiB/s

This is about a 29-31% throughput improvement on this workload.