ci: clarify LLGo dashboard benchmark groups

[zhouguangyuan0718]

Summary

• display each benchmark name once at the start of its build-mode group
• indent continuation build modes within the compact sticky label column
• add a full-width separator between benchmark groups in both matrices
• preserve accessible benchmark and mode labels on every row
• keep A/B selected-column outlines visible across group separators

Validation

• go test ./cmd/bent
• bash -n ci/llgo-size/report.sh
• git diff --check
• browser-tested with the current 41-run Pages history: 9 visible benchmark names for 54 rows in each matrix, 7px group separators, intact A/B highlighting, and no console warnings or errors

[fennoai]

Review summary

This is a small, focused UI change that groups the matrix rows by benchmark (name shown once per group, full-width separator before each subsequent benchmark). The implementation is clean and correct:

• HTML/attribute escaping is handled correctly — the new aria-label at app.js:312 wraps benchmarkName in escapeHtml, which encodes "/</>, preventing attribute breakout in the double-quoted value.
• The higher-specificity .benchmark-group-start ... > .selected-a/.selected-b rule (style.css:85-86) correctly layers the group separator on top of the existing accent side-borders rather than clobbering them.
• The README description (README.md:44-46) accurately matches the implemented behavior.

No blocking issues. One optional observation and one inline nit below.

Optional (pre-existing, not introduced here): rankFor (app.js:247, reached per cell via cellHtml) re-derives the full per-(benchmark, measure) ranking — filter + new Set + sort — once for every config row, even though the result is identical across all configs of the same benchmark. This PR's new outer-benchmarkName/inner-config loop shape now makes it natural to compute the ranking once per (column, benchmark) and reuse it across config rows. Not user-perceptible at current data scale; flagging only because the changed loop is the surrounding hot path.

[fennoai]

Minor (DRY): configIndex === 0 is tested three times in this loop (lines 310, 311, 316). Hoisting const isGroupStart = configIndex === 0; at the top of the callback and reusing it would remove the repetition and make the grouping intent self-documenting.