Using CI for relative benchmarking

Using CI to get absolute benchmark speeds is not feasible, but it may be possible to use it for relative benchmarks. https://dev.to/quansightlabs/is-github-actions-suitable-for-running-benchmarks-111h describes "relative" as between two commits, but we could also consider it between a mostly-fixed implementation (e.g., stdlib parsers) and the ones we're benchmarking. E.g., is it 1.5x, 3x, 100x slower than the stdlib parser, rather than 10ms, 20ms, 1s absolute times.

Also see https://github.com/marketplace/actions/continuous-benchmark