Allow for matrix-matrix and matrix-vector products with `KFACLinearOperator` and `KFACInverseLinearOperator` without converting to numpy

[runame]

Addresses #71, but only for the KFACLinearOperator and KFACInverseLinearOperator.

KFACLinearOperator/KFACInverseLinearOperator is arguably the only linear operator here that is likely to be used for preconditioned-gradient methods for large scale neural networks. Therefore, adding torch_matmat and torch_matvec methods seem like a simple solution to avoid unnecessary device transfers, which are a bottleneck for this use case. However, this doesn't address the issue in general.

[coveralls]

Pull Request Test Coverage Report for Build 8360156180

Details

• 52 of 58 (89.66%) changed or added relevant lines in 2 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.2%) to 89.367%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
curvlinops/inverse.py	16	17	94.12%
curvlinops/kfac.py	36	41	87.8%
<!--	Total:	52	58	89.66%	-->

Totals
Change from base Build 8250604944:	0.2%
Covered Lines:	1017
Relevant Lines:	1138

---

💛 - Coveralls