Mistral dropout

[winglian]

adds a gaussian dropout layer after the MLP layer and cleans up the model config so it's more easily modifiable from the YML

dropout for mistral can be added via

model_config:
  dropout_p: 0.001

[Mihaiii]

Any news on this? :) I'd love to try out the new config.

[winglian]

superseded by #858