[WIP] Integration of the latest `asdl` package

[aleximmer]

Enables regression and provides new interface for several methods that need to be synced (corresponds to #44). There are two key quantities that require more investigation and tests:

• [ ] regression empirical fisher seems to mismatch our backpack variant (currently failing test)
• [ ] regression fisher with sampling (stochastic=True) currently untested

[runame]

When this PR is ready and our backend also supports regression with ASDL, we can also change the default backend to AsdlGGN. This avoids issues with quite a few architectures, which are not implemented via a Squential, e.g. #92.

[adamxyang]

Hi, I'm also interested in the integration with the new ASDL. Does the current branch with the new ASDL work with classification tasks?

[adamxyang]

Looks like changing fisher_maker = get_fisher_maker(self.model, cfg, self.kfac_conv) to fisher_maker = get_fisher_maker(self.model, cfg) in asdl.py line 160 makes things work with the classification example.

[wiseodd]

Yes, that hotfix should work without any side effects. Let us know if you found any bugs, though!

[aleximmer]

@wiseodd you built on these changes for #144, right? Then, let's close this PR. Do you know if the two issues I report above are resolved?

[wiseodd]

Yes, should be good. Both are already covered with passing tests in #144. Closing this.