feat(autogram): Add Transformer support

[ValerianRey]

• Refactor how used params are deduced from the module: we now combine the direct params and the indirectly used params, and we reuse the same code between both usages of this function
• Add special case for indirectly used params of MultiheadAttention
• Add WithMultiheadAttention, WithTransformer, and WithTransformerLarge tests
• Add WithTransformerLarge speed test
• Rename trainable to rg
• Add test_batched_non_batched_equivalence_2
• Update warnings about Transformers in the docstring of Engine

Otherwise we'd have to wait for pytorch/pytorch#126568 to be merged.

Tests seem to fail because of a user warning, need to filter it out.

TODO:

• Isolate dirty code out of the rest
• Filter out user warnings
• Prevent hook creation for out_proj => More generally, we would like to prevent hooking module when they do not use any rg param. It's very hard to know in general which params are used, so we would need to do that on a case by case basis. However, in the case of mha.out_proj, I don't think it's worth the maintaining effort because out_proj isn't even called anyway, so its hook is never fired.
• Make sure this will be future-compatible: if no out_proj module exists anymore (i.e. if the pytorch PR is merged), it should still work => I tried modifying pytorch code to simulate a potential future and it still works, so we're good. This is because we check hasattr("out_proj") now.
• Add MultiheadAttention test
• Adapt warning about Transformers: we need to tell people to disable dropout and maybe some other parameters, but that's basically it
• Add usage example with Transformers => Not sure to do that yet actually, because it's not fast enough yet.
• Test more Transformer parametrizations

[PierreQuinton]

If PyTorch was to merge it quickly we would still only be compatible with very late versions which is a bit problematic. I think this suggests finding a solution independently.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
src/torchjd/autogram/_engine.py	100.00% <ø> (ø)
src/torchjd/autogram/_module_hook_manager.py	100.00% <100.00%> (ø)
src/torchjd/autogram/_module_utils.py	100.00% <100.00%> (ø)
src/torchjd/autogram/_vjp.py	100.00% <100.00%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[ValerianRey]

@PierreQuinton need review and we can merge!