Avoid redundant normalization of parsed wire headers

[Johnny-Kao]

Summary

• build canonical Headers directly while decoding header lines from the wire
• share Content-Length / Transfer-Encoding semantic normalization with the existing public normalization path
• preserve the existing syntax-validation-before-semantic-validation ordering

Why

The receive path currently validates each wire header line, materializes a temporary (name, value) list, and then passes that list through normalize_and_validate(..., _parsed=True), which walks the headers again to lowercase names, normalize Content-Length / Transfer-Encoding, and construct the internal three-tuple representation.

This change keeps the existing trust boundary and protocol checks, but removes the intermediate pair representation and redundant traversal. The public Request / Response construction path is unchanged.

Correctness / safety

The parser still validates every header line before applying Content-Length / Transfer-Encoding semantic checks. A regression test locks in that ordering.

Local validation:

• 79 passed
• Black 23.3.0: pass
• isort 5.12.0: pass
• mypy 1.8.0 strict config: pass
• git diff --check: pass
• 50,008 request-level differential cases matched the previous parser exactly for accept/reject behavior, error hint/message, request metadata, raw header casing/order, and canonical header values

The change does not alter obs-fold behavior; it is independent of #198.

Performance

Interleaved A/B measurements in the same Python process were used to reduce scheduler/frequency noise.

On the local Apple Silicon environment:

• repository realistic server GET cycle: ~3.3% lower median time (27.45 us -> 26.54 us)
• request receive / Connection.next_event() path: ~5.6% lower median time
• response header reader: ~11.7% lower median time
• response-parser traced peak memory in the test workload: ~11.4 KB -> ~6.3 KB

The performance claim is intentionally limited to these parsing/benchmark workloads rather than h11 overall.

[Johnny-Kao]

First time contributing to h11 — this was a really interesting codebase to work on. The implementation is impressively solid, especially around the protocol and validation boundaries. Happy to adjust anything if needed.