perf(header,uri): faster value validation, URI parse/format, map inserts

[geeknoid]

Four independent, benchmark-validated micro-optimizations:

• HeaderValue validation (from_bytes / from_maybe_shared / to_str): replace the early-return byte loop with a branchless OR-accumulation so the scan auto-vectorizes. The valid case scans the whole slice either way; the error path is rare.
• URI path/query scan: replace the per-byte range match (~10 comparisons per byte) with two compile-time 256-entry classification tables. Stays a const fn (used by from_static); verified to classify all 256 bytes identically to the old match in both the path and query states.
• URI formatting: Uri and PathAndQuery Display impls use direct write_str calls instead of write!/format-args machinery for their string parts.
• HeaderMap: mark hash_elem_using #[inline] so it specializes into the insert path.

Adds a gungraun (callgrind, deterministic instruction counts) benchmark suite mirroring the existing benches, plus a criterion wallclock bench (opt_paths) for the affected operations.

Before/after, gungraun instruction counts (deterministic):

  HeaderValue::from_bytes  short          118 ->     98   -16.9%
  HeaderValue::from_bytes  long           950 ->    196   -79.4%
  HeaderValue::to_str      short          175 ->    125   -28.6%
  HeaderValue::to_str      long          1319 ->    217   -83.5%
  Uri parse  relative_medium             1348 ->   1009   -25.1%
  Uri parse  relative_query              1779 ->   1194   -32.9%
  Uri to_string  relative                 553 ->    448   -19.0%
  Uri to_string  relative_query           995 ->    795   -20.1%
  Uri to_string  absolute                1873 ->   1525   -18.6%
  HeaderMap insert_all_std              19914 ->  18850    -5.3%
  HeaderMap set_10_std (n10)             3343 ->   2784   -16.7%
  HeaderMap set_20_std (n20)             6221 ->   5102   -18.0%
  HeaderMap hn_hdrs_set_8_get_miss       3453 ->   3076   -10.9%
  HeaderMap insert_custom (n500)       525783 -> 519283    -1.2%

Before/after, criterion wallclock (median; shared VM, noisy under ~100ns):

  hv_from_bytes_long              89.4 ns ->  33.3 ns   -62.8%
  hv_to_str_long                  89.8 ns ->   9.2 ns   -89.8%
  hv_to_str_short                  7.2 ns ->   3.9 ns   -46.6%
  uri_parse_relative_medium      130.9 ns -> 112.3 ns   -14.2%
  uri_parse_relative_query       196.6 ns -> 177.7 ns    -9.6%
  uri_to_string_relative          34.2 ns ->  26.2 ns   -23.4%
  uri_to_string_relative_query   112.7 ns ->  93.2 ns   -17.3%
  uri_to_string_absolute         193.4 ns -> 171.4 ns   -11.4%
  hm_insert_10_std               300.7 ns -> 296.5 ns    -1.4%

(hv_from_bytes_short is within wallclock noise here at ~28 ns; its gungraun instruction count drops 16.9%.)