coordinate_matrix: Fixed group boundaries for small matrices, fixed some corner cases in SpMV

For matrices with less than 64 rows, the group boundaries were calculated incorrectly, thus producing wrong results.
Moreover, in some corner cases the last entry computed by a group may not have been written to the result vector.
Both issues are supposed to be corrected, tests pass with a recent Intel OpenCL SDK on a CPU.
Tests on GPUs still required.