Input dropout

[Ghostvv]

Proposed changes:

• add dropout before ffnn

Status (please check what you already did):

• added some tests for the functionality
• updated the documentation
• updated the changelog (please check changelog for instructions)
• reformat files using black (please check Readme for instructions)

[tabergma]

Let's add a changelog so that we can find this change later again.

[Ghostvv]

@dakshvar22 @tabergma do we want to merge this one?

[dakshvar22]

@Ghostvv As discussed, can you test it on a couple of datasets and configs before merging it?

[tabergma]

This also requires retraining of models, doesn't it? Should we increase the minimum model version number and add a note to the changelog entry? Or should we push it to master so that we release it with the next minor?

[Ghostvv]

this one doesn't because dropout is turned off during prediction so prediction graph is the same

[Ghostvv]

but maybe merging it into master is safer bet

[Ghostvv]

let's merge 1.9.x in a separate PR first: #5660

[Ghostvv]

I added a parameter use_dense_input_dropout (similar to sparse one) and set it to False by default in EmbeddingIntentClassifier just to avoid surprises

[dakshvar22]

Would require increasing the minimum compatible version right? Since we added a new layer by default, it would be a model breaking version I guess.

[Ghostvv]

it needs to increase anyhow, since master is not compatible with 1.9, but I don't think it requires here, since dropout doesn't affect prediction graph