Optimize inline closer indexing

[bobzhang]

Summary

• build the inline closer index during tokenization instead of doing a second Deque::retain pass
• store closer positions in append-ordered arrays and use binary search for first-closer lookup
• bypass full inline resolution for token streams that contain only newline tokens

Validation

• moon check
• moon test
• cd cmarkwrap && moon check
• cd cmarkwrap && moon test

Profiling

• Confirmed moon run --profile --target native --release works locally with macOS Time Profiler.
• Parser-only spec.md.repeat(4) benchmark, 20 runs: current branch mean ~10096 vs fresh origin/main mean ~10293 in the same session. The win is modest, but the patch removes a real extra token pass and set-based closer lookup from the inline parser.