Add simple MLP to models

[jdb78]

MLPs often perform very well. We can add a MLP that only works on the decoder input.

[blacksnail789521]

@jdb78
Hi, I want to ask why do you use decoder_cont and decoder_cat to be the input of the model?
I thought the definition of "data windowing" is to feed encoder_cont and encoder_cat into the model.

Thanks.

[kvyaswanth]

decoder_cat can always be empty array and tft still works; decoder_cont is time varying known reals attributed to the targets, I guess they employed it into the model as they help for better prediction

but some times when I have to send in a sequence of consecutive timeseries >> max encoder length it gives

"AssertionError: filters should not remove entries all entries - check encoder/decoder lengths and lags" at somepoint