SVD-operation on the GPU (#11878)

* Added GPU implementation of SVD, but without complex number support and it requires m>=n

* Removed debug logs

* Fixed SVD: V-matrix was not transposed

* Fixed SVD implementation, there was a memory corruption error. Also renamed it to the suffix .cu.cc for consistency.

* Fixed stupid left-over from debug statements

* fixed formatting with clang-format

* small improvements based on the comments

* further tiny improvemens

* Refactoring, now it should also support m<n, but tests still fail sometimes

* Added explicit gpu tests

* Formatting

* Added checks if DoTranspose was successfull

* It really is a bug in cuSolver that it always returns zeros for Vt if n==1, but my workaround does not always work yet

* Fixed cuSolver issue if one dimension of the input matrix is one

* fixed buildifier error

* improved readability

* Fixed a stupid error. This happens if you make changes without testing it

* Fixed a bug if n==1 and full_matrices=false

* Implemented changes requested from the Github comments, Moved n==1-fix kernels outside of the batch loop to compute them once over all batches.

* Cleanup tasks

* Added TODO what has to be done to support complex matrices

* fixed issue that the order of } (end namespace) and #endif (GOOGLE_CUDA) was mixed up. This breaks the code if CUDA is disabled