Efficient MR logic to create Vocabulary and X and y Matrices

The current MR logic employs 2 sets of mappers and reducers to create the vocabulary and use the indices from the vocabulary to assign indices and find the X and y matrices. It'd be amazing to perform the two steps using a single set of mapper and reducer if possible.

Certain constraints such as the lexicographical ordering of the vocabulary and usage of only 1 input/output file can be done away with to support the new implementation.