Integrate benchmarking

[johannahaffner]

This follows up on #128.

Two small things:

• the script stays - having the benchmarks output this would entail customising the fixture provided by pytest-benchmark, and plotting is also something one would do every few runs, rather than every run - in particular whenever one would like to look at a comparison between solvers or implementation with others.
• We'll continue to require benchmarking runs to specify pytest benchmarks --benchmark-only - if we'd put --benchmark-only into benchmarks/conftest.py then it would get collected when pytest is invoked and it would override our default --benchmark-skip.

And one weird thing: if I remove the OrderedDict, I still get results jumbled between solvers. But as you (@patrick-kidger) note, it should not be needed. Playing with it for a bit I can remove it in two of the three places, but the underlying logic is not apparent to me, nor am I happy to trust that this would work for a different number of solvers / problems / something.
I think the pragmatic thing to do here is to simply document the puzzling necessity of OrderedDict in this script. (Low-key itching to check whether this is an upstream thing...)

[johannahaffner]

Tests are failing due to JAX 0.8.2 being used with 3.12, and a different version of JAX being used with 3.10, for which JAX has recently dropped support.

[patrick-kidger]

I think this looks reasonable to me, with only one comment, can we target dev instead of main?

[johannahaffner]

Now failing checks because dev includes a module that triggers the error being addressed in patrick-kidger/lineax#184. Tests pass and benchmarks run locally (benchmarks are skipped in CI and no other tested code is added).