Avoid allocations for complex segmenter runs

[Xuanwo]

Changelog

Performance: Avoid temporary complex-run input allocations in the published word and line segmenters.

Summary

This changes the published complex-language hot path to slice the original iterator input instead of building a temporary String or Vec<u16> for each complex run.

This is not a str specialization. The old iterator now keeps the iterator positioned at the current and previous code point, and the complex handoff is implemented through an internal ComplexRunSegmenter abstraction over Y::IterAttr:

• Utf8 slices CharIndices::as_str() and uses byte offsets.
• PotentiallyIllFormedUtf8 slices Utf8CharIndices::as_slice() and uses byte offsets.
• Utf16 slices Utf16Indices::as_slice() and uses code unit offsets.

Potentially ill-formed UTF-8 still goes through the UTF-8 complex segmenter over bytes, so invalid byte sequences continue to be interpreted through Utf8CharIndices as replacement characters. UTF-16 continues to preserve unpaired surrogates in the rule iterator and now passes the original UTF-16 slice to complex segmentation.

The same pattern is applied to both published word and line segmenters. The neo segmenters already route complex runs through Y::IterAttr and are not changed here.

Performance

Baseline: origin/main at 4f4c9f1fa068b6e4e0331b385104a7b2872489c5 plus the benchmark code from draft PR #8141 at 044cc38edec9ae46d03db66e9f5016bcd85e5835. Current: this branch plus the same benchmark code. Criterion command:

cargo bench -p icu_segmenter --bench bench --features unstable -- 'Segmenter/(word|line)/segment_(str|utf8|utf16)/published_dictionary/thai_han' --sample-size 10 --warm-up-time 0.2 --measurement-time 0.5 --noplot

Allocation counts were measured with a temporary local complex_allocs example over thai_han and are per full segmentation call. The allocation probe was not included in this PR diff.

Case (thai_han)	Allocs before	Allocs after	Bytes before	Bytes after	Criterion before	Criterion after	Time delta
word segment_str	16	11	760	512	1.0788 us	1.0423 us	-3.4%
word segment_utf8	16	11	760	512	1.1390 us	1.1560 us	+1.5%
word segment_utf16	16	11	634	512	0.9750 us	0.9984 us	+2.4%
line segment_str	15	6	288	120	0.8382 us	0.6498 us	-22.5%
line segment_utf8	15	6	288	120	0.9069 us	0.8195 us	-9.6%
line segment_utf16	15	6	198	120	0.6945 us	0.6226 us	-10.4%

The word segment_utf8 and segment_utf16 thai_han timings are small regressions in this quick Criterion run, but the temporary complex-run allocation is removed for all measured encodings and the measured line paths improve.