Compile s3 --include/--exclude patterns once per transfer

[Adityaj0]

Split out of #10545 at @aemous's request — this is half 2 of 2. The other half is #10551 (deferring the docutils imports); the two are independent and touch disjoint files.

Filter._match_pattern ran pattern.replace('/', os.sep) for every file for every pattern, and fnmatch.fnmatch normcased both the path and the pattern on every call before its internal cache lookup. All of that is constant across a transfer.

Patterns are now separator-normalized and translated to a compiled regex once per source type, and the path is normcased once per file instead of once per pattern.

Profiling 50k files against 6 patterns showed only 0.19s of the original 1.68s was actual regex matching — the rest was repeated setup:

                          ncalls  tottime
_match_pattern            600000    0.388
fnmatch.fnmatch           600000    0.274
fnmatchcase               600000    0.202
re.Pattern.match          600000    0.192   <- the only necessary work
posixpath.normcase       1200000    0.166
str.replace               600000    0.057

End to end:

100k files, 2 patterns    3.30 -> 2.12 us/file   (-36%)
100k files, 6 patterns    9.94 -> 5.99 us/file   (-40%)

Worth being clear about the scope of that number: this is client-side filter evaluation only. On a real aws s3 sync the network dominates, so this matters most for large listings where most files are filtered out — it is not a 40% reduction in overall sync time.

self.patterns / self.dst_patterns keep their existing shape and contents, since tests and external callers read them.

Testing

• tests/unit/customizations/s3, tests/functional/s3: 996 passed.
• Behaviour equivalence was checked with a differential test running the old and new matching over 400 randomized pattern/path combinations (mixed local/s3 source types, spaces, case variation, glob metacharacters in filenames): 0 mismatches.
• Two tests added covering filter reuse across source types and stability across repeated calls. Unlike a bug fix, these pass against both the old and new implementation by design — they are refactor guards, not proof of the change.
• ruff output on the touched files is unchanged from baseline.

I did not run the full tests/functional suite locally.

---

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.