New class: MixedDataset allowing for arbitrary data types

[raamana]

Perhaps new class might NOT be necessary, if we let MLDataset.dtype be array of numpy.dtypes corresponding to each feature, and indicate it during init?

Or we can define a MixedDataset class which takes care of all the init and cfg, and makes it easy to handle fully generic mixed datatypes.

Gotta work it out, and test it.