perf: cibench workflow

[AaronO]

First pass at adding benchmarks in CI.
Covers matrix of key archs and SIMD extensions.
Outputs delta against main with critcmp (with 5% noise threshold).

Future improvements

• Run classifier on results to indicate regression/improvement/neutral (aggregate or on specific arch)
• Post summary (or details if regression) as PR comment (didn't add cause without summary would be noisy)
• Could save main baselines as assets (and just ref/download them) instead of recomputing them for each PR
  (comes at a cost of statefulness and some adhoc complexity but would reduce CI time cost)

Notes

• This adds CI (compute) time, but don't think it adds wall time since there are longer subtasks
• CI is still more noisy than controlled local env, but probably good enough for substantial perf changes and sanity-checking

[seanmonstar]

Thanks!