Use Optuna directly instead of via Ray Tune for tuning

[mdekstrand]

It will be easier to integrate with sensible reporting if we just use Optuna directly, instead of via Ray Tune, to tune hyperparameters.

We can still support using Ray for that parallelism, and possibly even keep a simple version of Ray Tune support around, but a lot of the support code is fussing around to get Ray to play nicely with logging, progress, etc., which will be simpler in the common case with Optuna.

However, Optuna does not directly support checkpoint resuming, so that may be sub-optimal.