We need benchmarks to guide optimization work and catch performance regressions.

At the very least we should be benchmarking:

• Parsing
• Type checking
• Code generation

There's probably no point in doing this until the language/syntax has reached some kind of stability, as each new addition would probably make previous benchmarks redundant.

github-action-benchmark supports Rust, so makes sense to use that initially. It also supports syncing benchmark results with a GitHub pages branch, which would be neat.

The Cargo example for github-action-benchmark uses libtest, so I reckon we start with that initially. We can then consider moving to criterion once everything's working and/or there is a need for it.