[ADD] Minimal linear operator interface for PyTorch

[f-dangel]

Long-term, I want to add native PyTorch support for linear operators in curvlinops to address inefficiencies like #71, but also to clearly separate PyTorch from SciPy so that it will be easier to tackle features like supporting distributed settings.

This PR is a first step towards this goal.

From an API perspective, I plan to keep the constructor of all existing linear operators identical. The only backward-incompatible change will be that the produced linear operator will be purely PyTorch. To obtain the old behaviour one has to call .to_scipy() after the constructor.

Old: H = HessianLinearOperator(...)
Planned new: H = HessianLinearOperator(...).to_scipy(dtype=...)

The PR defines a linear operator interface in PyTorch which allows easy export to SciPy linear operators.
Importantly, the interface can multiply onto vectors/matrices represented by single Tensors, or a List[Tensor], which is more common in PyTorch. It verifies the input and output formats and all methods that need to be implemented assume the (more natural) tensor list format.

The next steps will be:

• Define a base class CurvatureLinearOperator that replicates curvlinops._base._LinearOperator but inherits from our PyTorchLinearOperator, rather than scipy.sparse.linalg.LinearOperator.
• Migrate each supported linear operator to inherit from CurvatureLinearOperator. I already tried that for the Hessian and was able to migrate without breaking the tests. I will set up a separate PR to keep the diffs manageable
• Once all operators have been migrated (and probably we can get rid of a lot of boilerplate to check shapes, e.g. in KFAC), we can remove the current base class in curvlinops._base.

Let me know if this makes sense.

[coveralls]

Pull Request Test Coverage Report for Build 11423244838

Details

• 57 of 77 (74.03%) changed or added relevant lines in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage decreased (-0.7%) to 88.295%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
curvlinops/_torch_base.py	57	77	74.03%

Totals
Change from base Build 10967050112:	-0.7%
Covered Lines:	1403
Relevant Lines:	1589

---

💛 - Coveralls

[runame]

I haven't thought much about what functions have to implemented but this LGTM! This will definitely be a useful refactor and enable/simplify new functionality.