Cusparse mats

[mkstoyanov]

Another code dump:

• global sparse operations now use the cusparse library
• handles the logic of doing all work on the GPU without moving back to the CPU side for the global kronecker
• did not use the existing sparse method, wanted to avoid allocations and keep data-structures persistent
• removed the whole pinned memory thing

[mkstoyanov]

@quantumsteve Let's try to get this done before Wednesday, this is the same algorithm as before with the addition of cuSparse for the global portion of the work.

The next PR is much shorter and has lots of cleanup. Also, it removes the need for the old kronmult code.