[ADD] EKFAC

[runame]

Implements EKFAC (and its inverse) support (resolves #116).

I think we should at some point refactor KFACLinearOperator and KFACInverseLinearOperator to inherit from KroneckerProductLinearOperator and EigendecomposedKroneckerProductLinearOperator (or similar) classes since torch_matmat and other methods can be shared. Also, currently KFACInverseLinearOperator doesn't support trace, det, etc. properties which can also be shared. I created #126 for this.

[coveralls]

Pull Request Test Coverage Report for Build 10975103378

Details

• 207 of 210 (98.57%) changed or added relevant lines in 2 files are covered.
• 2 unchanged lines in 1 file lost coverage.
• Overall coverage increased (+0.5%) to 89.5%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
curvlinops/inverse.py	36	37	97.3%
curvlinops/kfac.py	171	173	98.84%
<!--	Total:	207	210	98.57%	-->

Files with Coverage Reduction	New Missed Lines	%
curvlinops/kfac.py	2	93.71%
<!--	Total:	2		-->

Totals
Change from base Build 10408891176:	0.5%
Covered Lines:	1449
Relevant Lines:	1619

---

💛 - Coveralls

[runame]

@f-dangel One thing that is not tested and that could be wrong is the per-example gradient computation when there is weight sharing.

[runame]

Will continue this PR in ~2 weeks.