CUDA partitioned matrix view

Converts BlockSparseMatrix into two instances of CudaSparseMatrix,
corresponding to left and right sub-matrix.

Values of submatrix E are always just copied as-is, and values of
submatrix F are copied if each row-block of F submatrix satisfies
at least one of the following conditions:
 - There is atmost one cell in row-block
 - Row block has height of 1 row
Otherwise, indices of values in CRS order corresponding to value indices
in block-sparse order are computed on-the-fly.

Change-Id: I14eee00c36ee74b6b83fc85927907641383abfc7