Large nnz issues

[psythefirst]

As #2 found out, it is possible that nnz overflows the range of unsigned int. In this case, long int seems necessary for CSC/CSR indexing.

Another related problem is limited GPU memory. Suppose we have a matrix with nnz=2^32. Each element requires at least 8B (a float for value and an unsigned int for indexing). This translates to 32GB in total, which already exceeds most GPU memory size.