Faster conversion from numpy to cupy

[saishashank85]

def num_to_cud(k):
k = torch.tensor(k, device='cuda')
torch.cuda.synchronize()
k = cp.asarray(k.contiguous())
cp.cuda.runtime.deviceSynchronize()
return k

[jacobtomlinson]

I don't think adding pytorch as a dependency here is a good idea. I'm curious why this is faster and whether you did any measurements of the differences