Add CI support for benchmarking

[brandonwillard]

This PR adds support for automated benchmarking via pytest-benchmarking and github-action-benchmark. pytest-benchmarking worked well enough in my other project (i.e. http://github.com/pythological/unification/), so I figured we should try it here as well.

N.B. This is needed for the new Numba and JAX benchmarks we're in the process of adding.

• Refactor existing benchmark tests to use the pytest-benchmarking fixtures.
  A couple of old Scan tests were converted to benchmarks, and Numba and JAX benchmarks were added for a logsumexp graph using different input sizes.

Closes #718

[codecov]

Codecov Report

Merging #1347 (6d2031d) into main (ae182f0) will increase coverage by 0.07%.
The diff coverage is 99.42%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1347      +/-   ##
==========================================
+ Coverage   74.27%   74.35%   +0.07%     
==========================================
  Files         175      177       +2     
  Lines       48887    49046     +159     
  Branches    10375    10379       +4     
==========================================
+ Hits        36312    36468     +156     
- Misses      10282    10285       +3     
  Partials     2293     2293              

Impacted Files	Coverage Δ
aesara/compile/mode.py	84.47% <ø> (ø)
aesara/tensor/random/rewriting/basic.py	94.16% <ø> (ø)
aesara/link/jax/dispatch/shape.py	88.46% <93.33%> (+1.97%)	⬆️
aesara/link/jax/dispatch/random.py	100.00% <100.00%> (+2.63%)	⬆️
aesara/tensor/random/rewriting/__init__.py	100.00% <100.00%> (ø)
aesara/tensor/random/rewriting/jax.py	100.00% <100.00%> (ø)
aesara/tensor/rewriting/math.py	86.02% <0.00%> (-0.20%)	⬇️

[rlouf]

Should we at least have a benchmark that covers each backend to monitor compilation time?

[brandonwillard]

Should we at least have a benchmark that covers each backend to monitor compilation time?

Yes, definitely. We need to start adding benchmarks for a lot of things. I'm planning on repurposing some of those old Theano tests (i.e. the ones that are actually just examples) into benchmark tests.

We'll need to do that across a few follow-up PRs, but, in the meantime, it looks like we'll need to merge this so we can more easily test some of its functionality (e.g. comparisons of saved perf. data and new changes in a PR, saving to a separate perf. stats. site, the GitHub comment functionality, etc.)

[brandonwillard]

I've added some settings that I hope will prevent benchmark-action from breaking on PRs, while still comparing benchmarks from main and PR branches and failing the workflow when they're past a threshold.

We might need to merge this just to see if it works on the push events. After that, we can open a test PR with a failing benchmark test and see if the PR alert/failure works.