Adding dropout support for recurrent and non-recurrent connections in LSTM cell, Librispeech 100hr recipe incorporating training with max perturbation and stochastic/cascade dropout.

[jb1999]

Code changes to incorporate capabilities described in https://arxiv.org/pdf/1707.00722.pdf.

In particular:

• Librispeech 100hr recipe/setup
• support and recipe for max perturbation approach to data augmentation
• support and recipe for several dropout variants (both step and sequence implementations of RNNDrop, NML & forward dropout)
• support and recipe for naive, stochastic and cascade dropout combination.