cublas sgemm parallelization

[coreylowman]

Use cudarc::driver::result::stream api and cublas::result::set_stream to parallelize sgemm operations for conv2d and 4d batched matmul

[coreylowman]

Might be useful to create some Stream object that is easy to create & forces null stream to wait for it to finish before continuing. Could add to cudarc

[coreylowman]

The latest cudarc added a CudaStream object that we can use for this:

let stream = self.dev.auto_joining_stream()?;

self.blas.set_stream(Some(&stream))?;
// call kernel
self.blas.set_stream(None)?;

self.dev.join_async(stream)?; // or you can just `drop(stream)`;