Upgraded CuBLAS to API V2.

- more consistant error checking
- one cublas handle per GPU
3 files changed