Support batch processing for deep neural networks

[beomyeol]

This pull request introduces batch processing for deep neural networks. A batch size that is used for training can be specified in the neural network configuration file. The changes involved in this PR are the followings.

• An input batch is given for training a neural network or inferring outputs.
• The signature of push for ParameterProvider is changed. The size of batch should be given.
• The return matrix for generateParameterGradient is the element-wise sum of parameter gradient matrices for each input instance in a batch.
• The messages that are used for reducing parameter gradients with group communication are changed from a list of layer parameters to a list of pairs of an integer (the size of batch) and a layer parameter (the element-wise sum of parameter gradients for an input batch).

This closes #141.

[jsjason]

@beomyeol Sorry for the late pass. I've left some comments for you. The sample dataset works well on both groupcomm and paramserver configurations.

[beomyeol]

@jsjason Thank you for your review. I've addressed your comments and left a comment for the type of reduce values. Please take another look :)

[jsjason]

@beomyeol I've done a second pass and left two comments. Thanks!

[beomyeol]

@jsjason I've addressed your comments. Please review the changes.

[jsjason]

Looks great, will merge.