fix: show uncertainty in benchmark comparisons

[zeitlinger]

Improve benchmark comparison reporting without changing benchmark execution or sample configuration.

• Display JMH 99.9% confidence intervals beside PR and base means.
• Distinguish overlapping intervals, below-threshold changes, and missing/invalid uncertainty.
• Keep the verdict explicitly conservative and heuristic, not a significance test.
• Add coverage for rendering, malformed data, thresholds, metadata mismatch, and the supplied overlapping example.

This is separate from #2468 and does not rerun benchmarks.