CUBLAS: Improve concurrency using device pointer mode

[maleadt]

By default, CUBLAS uses the "host" pointer mode for scalar reference arguments: https://docs.nvidia.com/cuda/cublas/#scalar-parameters

julia> mode = Ref{CUBLAS.cublasPointerMode_t}();

julia> CUBLAS.cublasGetPointerMode_v2(CUBLAS.handle(), mode)

julia> mode[]
CUBLAS_POINTER_MODE_HOST::cublasPointerMode_t = 0x00000000

This allows us to use simple Ref boxes for scalar inputs, but results in calls like cublasDot blocking in libcublas until the GPU is synchronized. We should instead set the pointer mode to device, and synchronize in Julia, so that other tasks get the opportunity to execute while waiting for the GPU to finish.

Going beyond this, we could add async flags to these APIs to optionally return a lazy scalar, or a 0d array.

cc @Jutho