[ADD] Support CrossEntropyLoss in KFAC

[f-dangel]

Closes #44.

As a test, I am considering a data set consisting of a single sample. In that case, the Fisher-blocks are exactly Kronecker-factored, because there is no sum over multiple data points. In the limit of infinite MC-samples, KFAC thus approaches the block-diagonal Fisher.

[github-actions]

Pull Request Test Coverage Report for Build 6695531317

• 10 of 11 (90.91%) changed or added relevant lines in 1 file are covered.
• 40 unchanged lines in 2 files lost coverage.
• Overall coverage decreased (-5.1%) to 87.247%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
curvlinops/kfac.py	10	11	90.91%

Files with Coverage Reduction	New Missed Lines	%
curvlinops/examples/utils.py	6	41.67%
curvlinops/fisher.py	34	28.85%

Totals
Change from base Build 6695320687:	-5.1%
Covered Lines:	691
Relevant Lines:	792

---

💛 - Coveralls

[runame]

Besides the potential minor performance improvement in my comment, LGTM.