CTC code: Add testing code for context-dependent CTC (CCTC) transition-model object, and numerous bug fixes

Adding language modeling code (and test code for it) that is to be used as a phone language model in CTC modeling.

Adding script to generate lattices with alternative pronunciations of training utterances (to be used in CTC)

added max-change related stats in NaturalGradientAffineComponent

Fixing bug in ivector/logistic-regression-test.cc

Modified virtual functions Scale and Add 

Fixing bug in ivector/logistic-regression-test.cc and removing trailing whitespace from other logistic regression files

Adding nnet2 function NnetComputationChunked

Added --scale option to nnet3-am-copy, useful for shrinkage of parame…

nnet3 lstm train script bug fix

add log-prob-per-frame in nnet3 training log

nnet3: Efficiency improvement to nnet3-combine so it doesn't repeated…

nnet3: Efficiency improvement to nnet3-combine so it doesn't repeatedly have to compile the computation.

nnet3, LSTM scripts: add momentum, and remove --update-per-minibatch …

Adding nnet2 function NnetComputationChunked and updating associated unit test and binary. This version of NnetComputation divides the feature matrix into smaller chunks, and performs the posterior computation chunk by chunk. This makes the computation suitable for the GPU when the feature matrices are large.

nnet3: fix nnet3-train to make sure destructor of nnet3-train is called before we write out the model

nnet3, LSTM scripts: add momentum, and remove --update-per-minibatch (no longer needed due to optimization that has the same effect).  Caution: default=0.9, so it's on by default in the LSTM scripts.

Updates to LSTM recipe

nnet3: adding new optimization method that reduces redundant matrix c…

disabled debug mode of run_tdnn.sh script in AMI

Minor changes to lstm recipe. Minor change to AMI recipe for JHU cluster