Avoid using vmap when parallel_chunk_size=1

[ValerianRey]

• Add a special case to Jac when parallel_chunk_size=1 + comment to explain that this is a fix to vmap
• Add tests of Jac with chunk_size=1.
• Add tests of backward and mtl_backward with parallel_chunk_size=1.
• Add test with cyclic tensor graph
• Add changelog entry

Note that even without the change to Jac, and even with CUDA, test_tensor_used_multiple_times succeeds, so the issue is really with CUDA + RNN rather than with CUDA + the graph of tensors being cyclical. Still, I think this test is interesting so I kept it.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Files with missing lines	Coverage Δ
src/torchjd/autojac/_transform/jac.py	91.83% <100.00%> (+1.36%)	⬆️