perf: avoid copying DATA payloads in parser

[anmonteiro]

Summary

Avoid allocating a fresh bigstring for every parsed HTTP/2 DATA frame.

Instead of using Angstrom.take_bigstring for DATA payloads, carry a (buffer, off, len) payload view through Frame.Data and write directly into the Body.Reader buffer.

Why

Angstrom.take_bigstring copies.

On the request/response body hot path, that meant one intermediate bigstring allocation per DATA frame before the body bytes were copied into the reader-side Faraday buffer.

This change removes that intermediate allocation while keeping header-block parsing unchanged.

Validation

• dune build --display=short @runtest-test_h2 @runtest-test_h2_client @runtest-test_h2_server
• local 1.5 GiB upload benchmark against the local Eio benchmark server

Observed benchmark delta on the local upload benchmark:

• before: 4.21s (364.40 MiB/s)
• after: 3.70s (414.99 MiB/s)