Create utility classes for model output

[amanchoudhri]

Closes #19

[Aramist]

Looks good. Only thing I would do is run train/eval/assess (really quickly, even one epoch is fine) on all the different output types and make sure there are no fatal bugs. If that checks out then I think it's good to merge

[amanchoudhri]

Current todos before merge:

• [x] Test training one epoch and validating models of each output type
  • [x] Each Gaussian parameterization
    • [x] Fixed variance
    • [x] Spherical
    • [x] Diagonal
    • [x] Full
  • [x] MDN with Gaussian and Uniform
  • [x] Point predictor
• [x] Check that gradients are calculated properly, especially w.r.t. the UniformOutput and GaussianOutputFixedVariance classes
• [x] Run assess.py on a model outputting a probabilistic output
• [x] ...and on a model outputting a point prediction to check that it doesn't try to calculate calibration
• [x] ...and on an ensemble of models with different output types