[ADD] Minimal prototype for KFAC

[f-dangel]

Progress on #42.

This adds a first prototype for a linear operator of KFAC. It has very limited support (e.g. only Linear layers, no weight sharing), but I think this is a good milestone to start off with (and to keep the git diffs manageable. The tests compare the per-layer GGN with KFAC for deep linear models with MSELoss for a large number of MC samples from the predictive distribution.

[github-actions]

Pull Request Test Coverage Report for Build 6671908035

• 91 of 106 (85.85%) changed or added relevant lines in 4 files are covered.
• 37 unchanged lines in 2 files lost coverage.
• Overall coverage decreased (-5.7%) to 87.229%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
curvlinops/examples/utils.py	0	3	0.0%
curvlinops/kfac.py	86	98	87.76%

Files with Coverage Reduction	New Missed Lines	%
curvlinops/examples/utils.py	3	41.67%
curvlinops/fisher.py	34	28.85%

Totals
Change from base Build 6549326356:	-5.7%
Covered Lines:	683
Relevant Lines:	783

---

💛 - Coveralls

[runame]

LGTM! I think this is a good starting point and we can start adding the missing cases/functionality. Maybe we can create issues of the missing cases and distribute them among the two of us?