Reduce allocations in remaining ecosystem comparators

[andrew]

PR #30 removes temporary allocations from SemVer, RubyGems, and NuGet comparison. The same scanner and stack-buffer approach should help the remaining scheme-specific comparators.

Measured at b8f1741 with Go 1.26.5 on darwin/arm64 using representative version pairs:

• PyPI: replace PEP 440 regexp captures and release/local splits, currently about 1.16 µs, 803 B, and 6 allocs/op
• Maven: scan separators and digit/letter transitions without builders and temporary slices, currently about 757 ns, 1,392 B, and 20 allocs/op
• ALPM: scan typed segments into stack-backed storage, currently about 487 ns, 768 B, and 10 allocs/op
• Conan: replace recursive split-based parsing with views or stack-backed items, currently about 702 ns, 824 B, and 19 allocs/op
• Gentoo and APK: scan underscore suffixes and dot-separated base components directly, currently about 180 ns, 160 B, and 4 allocs/op
• OpenSSL: replace SplitN core parsing with direct indexes, currently about 102 ns, 96 B, and 2 allocs/op
• IntDot: compare dot-separated numeric components without temporary slices, currently about 105 ns, 112 B, and 2 allocs/op

Each change should keep exported APIs and scheme ordering unchanged, add differential or parity coverage against current behavior, and include before/after -benchmem results. PyPI and Maven should use separate focused PRs because their ordering rules are more involved.