[REF] Replicate base class, but inherit from PyTorch linear operator

[f-dangel]

This PR introduces a class CurvatureLinearOperator, which is almost identical to the current base class _LinearOperator, but inherits from PyTorchLinearOperator.

The next steps will be to (i) migrate each linear operator to inherit from CurvatureLinearOperator, rather than _LinearOperator, (ii) delete _LinearOperator.

[coveralls]

Pull Request Test Coverage Report for Build 11667319095

Details

• 65 of 116 (56.03%) changed or added relevant lines in 3 files are covered.
• 56 unchanged lines in 2 files lost coverage.
• Overall coverage decreased (-5.2%) to 86.417%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
curvlinops/utils.py	5	8	62.5%
curvlinops/_torch_base.py	57	105	54.29%

Files with Coverage Reduction	New Missed Lines	%
curvlinops/examples/utils.py	10	31.25%
curvlinops/fisher.py	46	26.47%

Totals
Change from base Build 11523760748:	-5.2%
Covered Lines:	1476
Relevant Lines:	1708

---

💛 - Coveralls

[coveralls]

Pull Request Test Coverage Report for Build 11523822133

Details

• 16 of 110 (14.55%) changed or added relevant lines in 2 files are covered.
• 72 unchanged lines in 3 files lost coverage.
• Overall coverage decreased (-8.2%) to 83.421%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
curvlinops/utils.py	2	8	25.0%
curvlinops/_torch_base.py	14	102	13.73%

Files with Coverage Reduction	New Missed Lines	%
curvlinops/examples/utils.py	10	31.25%
curvlinops/_torch_base.py	16	40.56%
curvlinops/fisher.py	46	26.47%

Totals
Change from base Build 11523760748:	-8.2%
Covered Lines:	1424
Relevant Lines:	1707

---

💛 - Coveralls

[runame]

LGTM!

Would be good to add unit tests, at least of all public methods.

[f-dangel]

Added the following test: In this PR, CurvatureLinearOperator has two public functions: to_device, and gradient_and_loss. The to_device function will be deprecated soon, therefore I am not adding a test here (it is also called by the other linear operators in their deterministic checks, therefore tested indirectly). I added an explicit test which makes sure that the loss and gradient are correctly accumulated over a data set. This functionality was tested indirectly before by making sure the curvature matrix-vector products are correctly accumulated over a data set. Now there is an explicit test which only relies on the base class.