Fix the Cuda transfer ops and test them.  Also since the test will always fail unless initialized in a very specific way, mark it a potentially failing not to alarm users.

Add a dependency on libcompyte since we can't get access to the functions through the gpuarray module (dlopen is not global).

Add tests for GpuFromCuda and CudaFromGpu (which I could not try).
Add tests for GpuAlloc inspired by the tensor tests.
While there adapt the makeTester function in the tensor tests for easier testing of other ops.

Add previous workaround for complexity in the grad code and some reassuring comments about the presence of an error message.