Handle batch size in Hidden state initialization for train set.

[giridharmunagala]

Batch size in the last batch is handled for Evaluation dataset but not for train. Resulting in wrong dimension for Hidden state for word attention net.

https://github.com/uvipen/Hierarchical-attention-networks-pytorch/blob/b1ea9e0b7bc294364f213e42507a6fe9d502a044/train.py#L85