MDGPS Cleanup/Refactor

[wmontgomery4]

Hey @cbfinn, the 'harley_refactor' PR is pretty far behind at this point, and I have some other new things that I wanted to add. This PR shouldn't really change anything functionally, and includes several things from that other pull request (which I'll close shortly):

1) Fixes issue with policy_sample_mode (no more keep_samples conflicting with it). 2) Restrict linesearch to the log space only, using bisection with a min/max_eta (we already have an explicit min_eta and implicit max_eta, and the bisection works pretty quickly without any of the stability issues from the old version). 3) Cleans up MDGPS quite a bit, making it faster and the code more legible. 4) Fixes the error in mjc_mdgps_example.

There are some larger architectural changes too, that clean up a lot of code: 1) The policy_prior/policy_prior_gmm classes do the fitting now, kind of like how dynamics_lr_prior does the fitting (this is a bit clunky still, but I like it more than having the full _update_policy_fit method) 2) The dynamics prior uses gauss_fit_joint_prior instead of doing the computation itself.

The only real functional change is that MDGPS no longer uses the classic/global step sizes. Now I'm just averaging predicted/actual_impr across the conditions and using a single step size for everything. This makes more sense for clustering experiments, and it's nice to not have to worry about another hyperparameter, which didn't seem to make much of a difference anyways (I still have a step_rule hyperparam, but it should normally be laplace, I'm just using the mc estimates for some debugging with reacher).

I haven't included any of the clustering stuff in here, nor single GMMs (not sure if everyone can use those anyways), and I'm going to base future clustering stuff on this branch.

[cbfinn]

Other than the one comment, this looks good! It would be nice to squash your commits into one commit before merging.

[wmontgomery4]

Squashed the commits, ready to merge when you are.

[cbfinn]

Thanks Harley!