Dropout is never disabled during sample() for RNNs

[skinnider]

See discussion here about the new models that Vishu is introducing: #289 (comment)
The opposite issue applies to the RNNs, the dropout is never disabled (we never set model.eval() in the sample method). After fixing the new models, we should add a call to .eval() in the RNN sample method and then make a similar change to call .train() after sampling is done.