Reduce startup text-rendering fragmentation

[tsuyoshi-otake]

Context: The latest x64 Release startup trace spends about 30.9 ms of 33.3 ms visible-line painting inside 343 ExtTextOut calls. Scope: use the existing privacy-preserving aggregate non-block categories, safely extend existing render runs for verified non-combining Japanese punctuation, add classification and pixel-equivalence regression coverage, and run paired A/B startup benchmarks. Preserve font, color, selection, clipping, Unicode cluster, and fallback boundaries. Non-goals: DirectWrite migration, blind layout parallelization, and additional AVX-512 work.

[tsuyoshi-otake]

Local implementation and verification complete (not committed yet). On tools/startup-benchmark-sample.md, 10 fresh plus 10 existing-profile runs per build all succeeded with cleanup verified. Fresh median first-content paint-lines: 36.965 ms to 21.234 ms (-42.6%); ExtTextOut time: 34.644 ms to 18.989 ms (-45.2%); calls: 343 to 87 (-74.6%). Existing-profile medians improved by 41.7%, 43.5%, and 74.6% respectively. Non-block CJK punctuation fell from 126 to 0. Debug app/tests build passed; 17 focused tests passed, including a 0-pixel-difference GDI comparison of legacy fragmented output versus the new block output; no test/editor processes survived.