[ADD] Support for KFAC with type-2 Fisher

[runame]

Resolves #49.

Note: I can't immediately see anymore why the implementation for CrossEntropyLoss works like this, will have to think about it, but should be correct.

[github-actions]

Pull Request Test Coverage Report for Build 6802629667

• 35 of 37 (94.59%) changed or added relevant lines in 1 file are covered.
• 40 unchanged lines in 2 files lost coverage.
• Overall coverage decreased (-4.5%) to 88.033%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
curvlinops/kfac.py	35	37	94.59%

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 6801631981:	-4.5%
Covered Lines:	743
Relevant Lines:	844

---

💛 - Coveralls

[f-dangel]

I decided to start a section 'Internals' in the documentation which explains the tricky parts in math and also contains links to derivations. This might be useful for others who are interested in extending this library, or learning more about the details of each curvature approximation.

I have worked on this topic for quite a bit, so it would be great to get feedback if it is comprehensible.

You can take a look here.

[runame]

I really like the refactor in terms of backpropagating the square root loss Hessian and the idea of having an internals section in the docs! I left a few comments/questions.