Handle time_major in lstm and remove reshape from embedding

[sonu1-p]

This commit makes the following changes:

1. Embedding was unnecessarily reshaping input to a rank2 tensor. This
   is not required and causes model to diverge from what its actually
   supposed to do.
2. Lstm was not considering time_major param. No transpose of input and
   output is required if the input is already time major, i.e of shape
   [ seq_len, batch_size, input_size]
3. Simplify the handling of case when return sequences is set to true.
4. Use squeeze and unsqueeze instead of reshape for dim representing
   direction of lstm for simplicity.