Improve hyperparameter organization

[kmax12]

Right now all parameters specified at one level for the whole pipeline. We should break up parameters by step of the pipeline.

This could also be used to improve the output of describe_pipeline

[dsherry]

I want to add a couple more points here which are related to this issue. These just came up in @angela97lin's catboost PR:

1. We should allow the sub-library to specify intelligent defaults whenever possible. Example: catboost's loss_function and bootstrap_type are both chosen by the library based on certain heuristics. Unless we have a good reason to override those heuristics, we should make use of them, because otherwise we risk negatively impacting performance.
2. Question: do we need a distinction between a) hyperparameters which automl will search over and b) the list of hyperparameters supported by a particular pipeline? This would be required if we wanted to have automl tuners operate on different sets of hyperparameters, so I guess the real question is whether or not we want to support that. Perhaps that should simply be filed as a future feature.

[dsherry]

@jeremyliweishih is this a duplicate of #359 ?

[dsherry]

Closing in favor of #359 and the other tickets we've got tracking hyperparameter improvements