perf(http): parse buffered headers without whole-line strings

[mizchi]

Header lines already present in the transport buffer currently go through async read_until and a whole-line UTF-8 String before the field name and value are copied out. This adds per-header work even when no I/O is needed.

This change adds an internal ReaderBuffer::take_buffered_until helper and a synchronous ASCII header parser. Buffered ASCII lines decode only the name and value; incomplete lines and non-ASCII text retain the existing paths. The borrowed view is consumed before the next read can mutate its buffer. Request-line parsing, cookies, duplicate-header rules, and body framing remain unchanged.

I collected these measurements while benchmarking Mars, my web server implementation. See this report for the methodology, tradeoffs, reproduction commands, harness, and raw data.

• Native Mars, 64 additional request headers: server instructions/request 593,563 → 415,189 (−30.1%). Standard request headers: −5.8%.
• Parsing 64 headers with 40-byte values: requested allocation 62,654 → 28,130 bytes (−55%), median time 17.313 → 10.669 µs.
• With 7-byte input fragments, the fast-path check adds about 1% allocation and was 2.4% slower in the microbenchmark. The benefit is concentrated in already buffered lines.

Measurements use independent native release builds based on upstream 43e41261f99261f4030efbed1970388930229baa, five timing/three allocation rounds and three network rounds. These are short runs on a shared developer machine, not a production throughput claim. This PR is independent of #603.

Validation on macOS:

• 115 native release tests across HTTP, io, read-buffer, coroutine, and event-loop packages; 104 native debug HTTP/io tests; 52 JS HTTP/io tests.
• Workspace-wide moon check --deny-warn passes for native, JS, Wasm, and Wasm-GC; formatting and generated interfaces updated.
• Tests cover fully buffered and 1/7-byte fragmented input, Unicode, invalid UTF-8, duplicate framing headers, repeated header names, and pipelined requests.
• Mars native tests: 92/92 on baseline and this patch; Mars JS suite: 331/331.