[MINOR] JIT optimize LibMatrixBinCell

This commit move some of the code inside LibMatrixBincell around to
encourage jit compilation of some methods. In specific folloing methods
have been introduced.

- safeBinaryMvSparseRowVector
- fillZeroValuesEmpty
- fillZeroValuesDense
- fillZeroValuesSparse
- safeBinaryMMDenseDenseDensePM_Vec (Plus Multiply kernel vectorized)
- safeBinaryMMDenseDenseDensePM     (Plus Multiply kernel small input)
- safeBinaryMMDenseDenseDenseContiguous (This one makes a big difference)
- safeBinaryMMDenseDenseDenseGeneric

In specific the safeBinaryMMDenseDenseDenseContiguous,
safeBinaryMMDenseDenseDensePMm and safeBinaryMMDenseDenseDensePM_Vec
improve the performance by much.

In LM_cg the performance:
Stats output:

 +*  3.123   3000 (Before)
 +*  1.991   3000 (After)

 +   1.125   2021 (Before)
 +   0.703   2015 (After)

This is training on Criteo 100k rows.