Add Curvlinops backend & add default functorch implementations of many curvature quantities

[wiseodd]

Changelog:

1. Add Curvlinops backend.
2. Add default functorch implementation for many quantities: (in CurvatureInterface itself)
   1. Jacobians
   2. Batch gradients
   3. Diag GGN (both stochastic and exact) and diag EF
   4. Full GGN (both stochastic and exact) and full EF

Rationale for (1): BackPACK & ASDL are not actively supported anymore and are hard to extend. Note that we use Curvlinops mainly for KFACs.

Rationale for (2): functorch is a first-class PyTorch module => has the highest compatibility to any PyTorch models => futureproofing laplace-torch, less dependence to third-party backends

Together: They make laplace-torch very flexible and compatible with most architectures out there (and in the future).

With this PR, Laplace(backend=CurvlinopsBackend) supports:

1. Diag-GGN-exact, diag-GGN-MC, diag-EF for both classification and regression
2. Kron-GGN-exact, kron-GGN-MC, kron-EF for both classification and regression
3. Full-GGN-exact, full-GGN-MC, full-EF, full-Hessian-exact for both classification and regression

Note that previous backends only support some subsets of them.

I'd like to ask @runame to specifically review this since he's involved in Curvlinops. Specifically, for the K-FAC backend proposed here.

[runame]

I reviewed the curvlinops part for now and pushed fixes directly.

I noticed that our tests are not systematic enough and that we are currently missing a few important cases (e.g., KFAC(-expand) is not the same with backpack and curvlinops for Conv2d modules). Since this branch still uses the old ASDL version, I think we should fix the tests of the curvature backends after this PR and #144 are merged.

[wiseodd]

@aleximmer please give a final check and feel free to merge!

[aleximmer]

Looks overall good to me. Please have a look at the two comments and let me know what you think. After that it's good to go from my side.

Thanks for taking on the effort and including curvlinops!

[aleximmer]

Sorry, forgot to add this comment in the previous review.