Add benchmark workflow

[sinsoku]

Motivation / Background

We want to prevent regressions by running TypeProf on real-world Ruby code.

Additionally, we want to measure improvements (or deteriorations) in performance and type inference accuracy.

Detail

Run TypeProf against TypeProf itself, Optcarrot, and Redmine on each
push to master, and publish the following metrics as charts on GitHub
Pages:

• elapsed time (seconds)
• type coverage per slot (% of argument/return slots inferred)
• type coverage per method (% of fully-typed methods)

Example (running on my fork):

• https://sinsoku.github.io/typeprof/

Setup

Please set Settings → Pages → Source to GitHub Actions (once,
admin only) so the deploy job can publish the dashboard.

Resolved questions

1. Remove tool/dog_bench.rb?
   It overlaps with tool/benchmarks/typeprof.rb, but might still be
   useful for ad-hoc profiling.
   -> Keep it; still useful for ad-hoc profiling.

2. Run benchmarks on each PR?
   Catches regressions earlier, but each run takes a few minutes.
   -> Added as a CI test (not a benchmark) against real-world projects.

[sinsoku]

@mame
What do you think of the overall direction? Want to make sure I'm on the right track.

[mame]

This looks really nice! I'd like to give it a try.

A few thoughts on the open questions:

1. tool/dog_bench.rb: I'd like to keep it for now. I do use it for ad-hoc profiling sometimes.
2. Running on each PR: I think this would be valuable, but more as a test than as a benchmark. It would help us catch cases where a change to TypeProf accidentally breaks it on Redmine or Optcarrot.

[sinsoku]

@mame
I've resolved both points.

1. Kept tool/dog_bench.rb as-is.
2. Run the benchmark scripts in CI.

The CI uses tool/benchmarks/typeprof.rb instead of tool/dog_bench.rb, because pinning the analyzed version keeps the input stable and makes regressions easier to detect.