When Uri\Rfc3986\Uri::parse() produces a URI already in canonical form
(the common case: http/https URLs with no uppercase host, no
percent-encoding in unreserved ranges, no ".." path segments),
get_normalized_uri() no longer deep-copies the parsed struct and runs
a full normalization pass. It calls uriNormalizeSyntaxMaskRequiredExA
once to compute the dirty mask; a zero mask means we alias the raw
uri. The struct caches the dirty mask, so multiple non-raw reads on
the same instance only run the scan once.

Fallback: when the mask is nonzero, we copy and normalize as before,
but only for the flagged components (uriNormalizeSyntaxExMmA(...,
dirty_mask, ...) instead of (..., -1, ...)).

Measurements on a 17-URL mix with a realistic parse-and-read workload
(10 runs of 1.7M parses each, CPU pinned via taskset, same-session
stash-pop A/B so both builds share machine state):

                        baseline mean    optimized mean    delta
    parse only         0.3992s (4.26M/s)  0.4083s (4.16M/s)  noise
    parse + 1 read     0.6687s (2.54M/s)  0.5464s (3.11M/s)  -18.3%
    parse + 7 reads    0.8510s (2.00M/s)  0.7305s (2.33M/s)  -14.2%

The "parse + 1 read" row isolates the first-read cost where this
change lands. The "parse + 7 reads" row shows the amortized effect
under a realistic user pattern: the first getter pays the reduced
normalization cost, and the remaining six getters hit the cached
normalized uri and cost the same as before.

hyperfine cross-check on the whole benchmark script, 15 runs each:

    baseline   20.471 s +/- 1.052 s  [19.535 .. 22.985]
    optimized  17.240 s +/- 0.540 s  [16.556 .. 18.190]
    optimized runs 1.19 +/- 0.07 times faster.

All 309 tests in ext/uri/tests pass. I checked that URIs needing
normalization (http://EXAMPLE.com/A/%2e%2e/c resolving to /c) still
hit the full normalize path through the nonzero dirty mask.