Emulator descriptors

[marcosfelt]

The ExperimentalEmulator did not allow using descriptors previously with categorical variables. Now a list of categorical variables names can be passed to descriptors_features, and the associated descriptors from those variables will be used in the transformation.

This also fixes #99, by automatically converting spare numpy arrays to dense before conversion to a pytorch tensor.

To-Do:

• Retrain Baumgartner w/without descriptors
• Add normalization to the descriptors transform
• Create two json files - one include_cost and the other not
• Selectively choose right model_name based on using descriptors and include_cost
• Update tests to test if include_descriptors works properly