perf(ci): parallelize browser regressions and report phase timings

[xiaoiver]

Browser regression currently runs Chromium and WebKit sequentially with one worker, taking 38–46 minutes in recent successful jobs. Run two Chromium shards and two duration-balanced WebKit groups (Lottie / all other suites) on independent runners. Preserve one worker per runner and the existing WebKit process isolation.

Keep browser-regression as the aggregate required check. It only succeeds when static checks and every browser group succeed; failed, cancelled and skipped prerequisites cannot pass the gate. Cancel superseded runs for the same PR/ref and retain per-group reports and failure artifacts for seven days.

Add a timing reporter with setup/body/teardown breakdowns, slow-file summaries and machine-readable results. Tooling tests discover the real Playwright suites to verify that the workflow groups cover every case exactly once: both Chromium shards and both WebKit groups, including the newly merged Brush cases. Local commands still run the full suites by default. No test assertions, retries or per-test timeouts are relaxed.

Collaboration replicas now initialize only the canvas they use. The main page and all canvas-isolation/lifecycle scenarios retain both canvases. This avoids two unused WebGL contexts across the additional tabs; locally the Yjs case improved from exceeding its 30-second limit to ~20 seconds.

Validation:

• 18 tooling tests passed, including exact test partition coverage, merge-gate failure/cancellation/skipping, and timing aggregation.
• Browser TypeScript, root ESLint, Markdown lint and diff checks passed.
• All 32 WebKit Lottie cases, 7 Chromium Mermaid cases, and the WebKit D2 documentation case passed with the actual CI reporter and summary output.
• All 6 Chromium lifecycle cases passed, including Yjs and Loro edits, undo/redo, deletion, late joining, resource cleanup and canvas isolation.
• The complete browser suites will run in this PR's CI.

The previous CI log puts WebKit Lottie at ~11.6 minutes and the remaining WebKit cases at a similar cost. The target is approximately 15 minutes overall with available runners; actual CI timing is not yet measured.