in test_nnet use the new verbose flag and allow to use only the cpu version and allow to don't pickle.

If the THEANO_GPU env var is set to a number, we will use this card number.

some small fix to make it work when the end don't have CUDA_ROOT, but that the other env variables are set correctly.

- pickling and unpickling them when they are in graph keys.

-  adding a Signature type.

Mainly:
- improved GpuSum.__str__
- fixed problem in make_node of IncSubtensor and Subtensor
- added PyErr_Format()s to failure-handling code of GpuSum

But it depends on some very sketchy code in cuda_ndarray, so it will need
further work.