perf: fast path for normalized absolute special URLs

[crowlbot]

Summary

Adds a single-pass fast path for parsing already-normalized absolute URLs, modeled on ada-url's try_parse_simple_absolute.

Today every Url::parse call runs the full WHATWG state machine (char-by-char iteration with tab/newline filtering, percent-encode set lookups, IDNA, etc.). The overwhelming majority of real-world URLs are already normalized. This PR adds two sibling byte-scan fast paths that construct the Url directly for them, deferring anything that needs normalization/encoding to the existing parser so observable behavior is unchanged:

• try_fast_parse — special schemes (http/https/ws/wss/ftp): clean ASCII host (mixed-case allowed, lower-cased like IDNA would), optional canonical port, verbatim path/query/fragment (percent-escapes in the path are kept verbatim; only dot-segments defer). Because it only accepts complete absolute URLs (scheme://…), which the WHATWG algorithm resolves without the base, it also serves Url::join with an absolute argument.
• try_fast_parse_nonspecial — non-special schemes with an authority (redis://, postgres://, web+demo://, …), which serialize verbatim (opaque host, no lower-casing/IDNA/IPv4, no default port, empty path stays empty).

Correctness

• In debug builds, every fast-path hit is cross-checked field-by-field against the general parser using the same base URL (debug_check_fast_parse). The test suite runs in debug, so the full WPT urltestdata corpus (thousands of URLs, many with base URLs, mixed-case hosts, and non-special schemes) validates byte-for-byte equivalence, including base-independence. This guard caught several subtle cases during development (xn-- punycode labels, IPv4-looking hosts, default/leading-zero ports), all now deferred.
• New test_fast_path_parse covers accepted (special, ports, mixed-case, non-special), deferred, and join-vs-parse inputs explicitly (also runs check_invariants).
• All existing url tests pass (unit + wpt).

Benchmarks

cargo bench -p url --bench parse_url (median of runs; unicode_*/punycode_* don't hit the fast path and are unchanged):

bench	before	after	change
short	126 ns	70 ns	−44%
long	149 ns	80 ns	−46%
plain	112 ns	71 ns	−37%
port	149 ns	58 ns	−61%
hyphen	153 ns	73 ns	−52%
leading_digit	161 ns	72 ns	−55%
fragment	160 ns	84 ns	−48%
uppercase_host	194 ns	82 ns	−58%
nonspecial_scheme	125 ns	51 ns	−59%
encoded_path	214 ns	103 ns	−52%
ipv4_host	286 ns	81 ns	−72%
join_absolute	141 ns	84 ns	−40%

[crowlbot]

Extended the fast path to also cover a canonical host:port authority (all-digits, in-range, no leading zeros, non-default port). Ports that need normalization (default-port stripping, leading zeros) still defer to the general parser, and the debug cross-check + WPT suite continue to validate byte-for-byte equivalence.

New benchmark:

bench	before	after	change
port (https://example.com:8080)	149 ns	58 ns	−61%

[crowlbot]

Relaxed the base_url.is_none() gate so the fast path also serves Url::join/relative-parse with an absolute argument (e.g. resolving an absolute href against a page's base URL — a very common operation).

try_fast_parse only accepts complete absolute special URLs (scheme://…), which the WHATWG algorithm resolves without consulting the base, so base.join(abs) is guaranteed identical to parsing abs standalone. The debug cross-check now re-parses through the general parser with the same base, so the WPT urltestdata corpus (which is full of base + input combinations) validates base-independence.

New benchmark:

bench	before	after	change
join_absolute	141 ns	84 ns	−40%

[crowlbot]

Extended the fast path to accept mixed-case ASCII hosts (e.g. https://GitHub.com from user input). For a host made only of ASCII [A-Za-z0-9.-], IDNA's only effect is ASCII lower-casing, so the fast path lower-cases the host itself (length-preserving, so all byte offsets stay valid); the xn-- and IPv4 checks run on the lower-cased form. The debug cross-check + WPT urltestdata (full of mixed-case host cases) validate equivalence.

New benchmark:

bench	before	after	change
uppercase_host (https://Example.COM/bench)	194 ns	82 ns	−58%

[crowlbot]

Added a sibling fast path for normalized absolute non-special URLs with an authority (redis://, postgres://, mongodb://, web+demo://, custom app schemes) — a common shape in backend/config parsing. These serialize verbatim (opaque host, no lower-casing, no IDNA/IPv4, no default port, empty path stays empty), so the fast path just validates and copies. Kept as a separate function (try_fast_parse_nonspecial) so the special-scheme path is untouched. Anything needing normalization/encoding defers to the general parser; the debug cross-check + WPT validate equivalence.

New benchmark:

bench	before	after	change
nonspecial_scheme (redis://localhost:6379/0)	125 ns	51 ns	−59%

[crowlbot]

Extended the fast path to accept percent-escapes in the path (e.g. /wiki/C%2B%2B, /a%20b) — previously any % in the path forced the slow path, despite these being extremely common. The URL parser leaves % and its following bytes verbatim in the path; the only path normalization is dot-segments. So % is now a verbatim path byte and the dot-segment guard is expanded from ./.. to the exact set the parser rewrites (including the %2e/%2E escape forms — is_dot_path_segment). Applies to both special and non-special path scans; the debug cross-check + WPT (rich in %2e/dot-segment cases) validate equivalence.

New benchmark:

bench	before	after	change
encoded_path (.../wiki/C%2B%2B%20programming)	214 ns	103 ns	−52%

[crowlbot]

Extended the fast path to accept canonical dotted-decimal IPv4 hosts (https://192.168.1.1/, http://10.0.0.255/) — common in dev/internal/IoT contexts and previously always deferred (the host "ends in a number", ~286 ns). An already-canonical address (four base-10 octets 0–255, no leading zeros) serializes verbatim and maps directly to HostInternal::Ipv4; parse_canonical_ipv4 recognizes exactly that form and defers everything else (leading zeros, hex/octal, <4 parts) to the general parser. Debug cross-check + WPT (extensive IPv4 coverage) validate equivalence.

New benchmark:

bench	before	after	change
ipv4_host (https://192.168.1.1/status)	286 ns	81 ns	−72%