[Maintenance] testing: Per-PR perf regression detection (BDN delta vs base branch)

[Chris-Wolfgang]

Scope

#184 / P2 (benchmark gh-pages chart) graphs the BDN trend on every push to main — a backward-looking signal that surfaces regressions after they merge. The deferred work is the forward-looking signal: compare the PR's BDN run against its own base branch and post a delta-table comment on the PR.

A PR that doubles allocations on a hot path should be caught before merge, not after.

Acceptance criteria

• A pr-benchmarks.yaml workflow runs the benchmark suite on the PR's HEAD and on base...HEAD's parent, captures both BDN JSON outputs, computes the per-benchmark delta (time / allocations), and comments on the PR with a markdown table.
• The comment is replaced on each push to the PR (not appended) so the latest delta is always at the top.
• The workflow fails the PR if any delta exceeds a configurable threshold (e.g. 20% slower or 50% more allocations) AND the PR didn't tag itself perf-impact-acknowledged.
• The workflow only runs when src/ or benchmarks/ files change (path filter).

Notes

• Tier-1 follow-up. Distinct from fix: unsubscribe Elapsed on Dispose for all doubles #184 (P2 — graphs trend); this is per-PR gating.
• Originated as part of the IAsyncEnumerable-Extensions thorough-review pass; tracked fleet-wide.
• Board: https://github.com/users/Chris-Wolfgang/projects/6