Describe the bug
I've been trying to understand the differences between BlockRNNModel and RNNModel; in particular, why one only supports future covariates, one only supports past covariates, and neither support static covariates. This is still unclear to me, however it's not the reason for this issue.

While I was looking at the source, I noticed that in BlockRNNModel, an MLP is being created which is just a stack of linear layers with no activations in between, which seems like a mistake.

Expected behavior
There should be an nonlinear activation function between the linear layers, otherwise it is the same as just having a single linear layer. An argument to BlockRNNModel should be provided that allows the user to choose this activation function.