Add mlp_encoder_model (MLPs are used to process input data)

[lgray]

This model should probably only be used with "slim" 3 output variable training (regress x, y, beta) and "diagonal" 8 variable output training (x,y,alpha,beta + errors).

[lgray]

@GiuseppeDiGuglielmo FYI

[lgray]

@davidgjiang @ArghyaDas112358 please have a look here.

For this model specifically we'll need two different loss configurations:

• "diagonal": this is a mixture density network where the covariance matrix is only the diagonal, so just the diagonal of the cholesky decomposition too
• "slim": this is a standard MSE regression to the x, y, cotbeta so no need to initialize a gaussian or anything

I am not confident that this model can handle the full covariance training.

Please feel free to add to this PR with your edits!

[lgray]

@GiuseppeDiGuglielmo Please give a try to synthesis to see how it goes for 2 timeslices!

[lgray]

@davidgjiang @ArghyaDas112358 I had a typo and forgot to actually quantize the MLP encoder. Please update and retrain!