perf: return relaxed header values to the fast scanner

[mcollina]

Fixes #881.

The relaxed header-value matcher previously stayed in its own loop, duplicating the ordinary header scanner (including vectorized blocks) in the generated state machine. After accepting a relaxed byte, return to the ordinary header_value scanner. Keep invalid bytes on an error path and cover consecutive relaxed bytes and a bare CR following one.

Local -O3 standalone benchmarks (4 and 32 headers) improved with clang; gcc remains layout-sensitive. For example, one pinned-core 32-header interleaved run measured ~1,470–1,590 ns/request before and ~1,290–1,350 ns/request after with clang. These numbers are machine/compiler dependent.

Validation: npm test (276 passing), npm run lint, npx tsc --noEmit -p tsconfig.json, git diff --check.

[mcollina]

@pimterry PTAL