Do full SVD on GPU (nvidia svd cu stuff is mostly on CPU using lapack)

[pseudotensor]

I thought cuda9 would improve this, but seems they didn't improve this aspect. Even MAGMA package has the same issue. I understand some stuff is inefficient for the GPU, but maybe they only tested for small data when assuming to do significant work on the CPU.

It's still true that CPU is slower, because what work is done on the GPU is done very efficiently, but for large matrices the work on the CPU takes way too long to be useful.