Fix crash in new cuda tensor with numpy array

[vedanuj]

Fixes segfault in #5821 . Releasing the Global Interpreter Lock after maybe_initialize_cuda.

Test ::

In [3]: data = np.array([1.0])

In [4]: torch.tensor(data, dtype=torch.float32)
Out[4]:
 1
[torch.FloatTensor of size (1,)]

In [5]: torch.tensor(data, dtype=torch.cuda.float32)
Out[5]:

 1
[torch.cuda.FloatTensor of size (1,) (GPU 0)]

Please review @colesbury @gchanan

[soumith]

lgtm, thanks vedanuj!