Fix/query encode semi - #288

Closes upstream issues unjs#233, unjs#240, unjs#301, unjs#302, unjs#304 in a single pass, and
verified against the WHATWG URL spec (\u00a75) + native URLSearchParams output.

Every code point outside `[A-Za-z0-9*-._]` is now percent-encoded by
`encodeQueryValue` / `encodeQueryKey`. Direct char-by-char parity check
against `new URLSearchParams([['k', c]]).toString()`: 33/33 match.

Concrete deltas from prior behavior:
  ^ ->  %5E  (was raw, restored from encodeURI output; issue unjs#304)
  ` -> %60  (was raw, restored from encodeURI output; issue unjs#302)
  | ->  %7C  (was raw, restored via encode() pipe rewrite; issue unjs#233)
  ; ->  %3B  (issue unjs#240)
  ? ->  %3F  (issue unjs#301)
  ! ->  %21    ' -> %27    ( -> %28    ) -> %29
  ~ ->  %7E    , -> %2C    : -> %3A    @ -> %40
  = ->  %3D  (in values too, matching URLSearchParams; encodeQueryKey no
              longer needs a separate `=` pass, so it becomes an alias).

Community consensus: 11 open upstream PRs (unjs#279, unjs#288, unjs#303, unjs#305, unjs#310,
unjs#318, unjs#324, unjs#327, unjs#328, unjs#329, unjs#354) all propose subsets of this fix.
This commit lands the full set at once.

Path/hash encoding is UNCHANGED \u2014 the WHATWG fragment percent-encode set
excludes `^{}` and `|` on purpose (fragment allows them raw), and encode()
retains its pipe-restore for path/hash consumers.

Tests: 991 pass + 65 xfail. Added WHATWG-URLSearchParams parity regression
test that walks every contested char and asserts native-parity.

…paque schemes)

Encoding — WHATWG application/x-www-form-urlencoded compliance
- encodeQueryValue/encodeQueryKey: byte-for-byte parity with
  URLSearchParams.toString(); |, `, ^, @, :, ,, ;, =, ? all encoded
  (adopts consensus from 11 open upstream PRs: unjs#279 unjs#288 unjs#303 unjs#305
  unjs#310 unjs#318 unjs#324 unjs#327 unjs#328 unjs#329 unjs#354)
- Single-pass replacer; %20 → + rewrite after encodeURI, before map

parseQuery — correctness + security
- Single-pass charCode scanner (PR unjs#331 @saripovdenis, ~40% faster)
- Empty-key preservation: =value → { "": "value" } (PR unjs#355 @spokodev;
  was silently losing the value with the old /([^=]+)=?(.*)/ regex)
- Proto-pollution guard: __proto__ / constructor / prototype blocked
  (PR unjs#289 @pi0; extended to cover prototype as well)
- stringifyQuery: single-pass builder (PR unjs#333 @saripovdenis)

parseURL — opaque-scheme URIs (RFC 3986 §3)
- mailto:, tel:, urn:, sms: — scheme NOT followed by // → opaque path
  surfaced in .pathname; query/fragment parsed from the opaque tail
- data:, blob: round-trip correctly (tested against WPT urltestdata.json)

parseAuth — RFC 3986 §3.2.1
- Split on FIRST colon only; subsequent colons belong to the password
  (was splitting on all colons, losing interior password chars)

parseHost — IPv6
- Brackets retained on hostname to match WHATWG URL.hostname contract
- Malformed unclosed bracket returns input verbatim (no silent truncation)
- Non-numeric port suffix detected (was silently truncated)

cleanDoubleSlashes — query/fragment protection
- Double-slash collapse no longer touches query or fragment sections

hasProtocol — single-char scheme rejection
- Blocks Windows drive letters (C:) and bare-digit prefixes

withBase / withoutBase — fragment-boundary fix
- Fragment on input (#hash) no longer defeats 'base already present'
  check (was producing doubled base path on fragment-suffixed inputs)

withFragment — empty hash strips existing fragment

WPT urltestdata.json ratchet
- 100-case special-scheme subset; 65 known-divergent cases run via
  it.fails (same count as upstream); ratchet trips if a fix lands