Remove loss_average argument of KFACLinearOperator

[runame]

Since we added the num_per_example_loss_terms recently, we could remove the loss_average argument. I think there is no use case where it adds any additional benefit, and removing it will simplify the user experience significantly and it avoids user errors -- when the user doesn't pass any value for num_per_example_loss_terms it will still lead to the correct result, there will just be the overhead of one dataset pass.

Moreover, I created Enums for the string valued arguments in KFACLinearOperator (FisherType and KFACType) and added a custom TypeVar for the input/ouput of torch_matmat/torch_matvec to support static type checkers (this was a request by @BrunoKM).

[coveralls]

Pull Request Test Coverage Report for Build 9288405107

Details

• 31 of 31 (100.0%) changed or added relevant lines in 2 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.2%) to 88.948%

---

Totals
Change from base Build 9214967831:	0.2%
Covered Lines:	1328
Relevant Lines:	1493

---

💛 - Coveralls

[f-dangel]

LGTM!