Track benchmarks over time in CI #298

🌀 Claude: The ML demos' numbers are measured by hand and posted on #285, with the CPU and versions. Nothing records them over time, so a regression in xarray-sql, ddx, einfold or DataFusion shows up only when someone measures again.

Proposal: track benchmarks over time in CI, starting small:

• pytest-benchmark for the Python benchmarks, including the ML demos at fixed sizes. Each step's time and its error against JAX are already in _bench.py's results table and CSV.
• criterion.rs where the Rust code needs it.
• github-action-benchmark to store the history and flag a regression on a PR.

Hosted runners vary from run to run, so the thresholds need to allow for that noise. Alternatives: codspeed.io, which counts instructions rather than measuring time, and bencher.dev. Both are free for open source today, but carry the risk of a SaaS changing its terms. asv is a reference for what a full history and dashboard could look like.

Raised in review of #269. The performance targets themselves are tracked on #285.