Fix QA pipeline on Windows

[sgugger]

What does this PR do?

As reported on the forum, there is a problem in the current pipeline on Windows. The problem's root is that numpy int arrays have a different default on Linux and Windows, the current snippet:

import numpy as np
x = np.array([1, 2, 3])
x.dtype

will print dtype('int64') on Linux/MacOS but dtype('int32') on Windows. So this means that just doing torch.tensor(some_numpy_array) may result in a tensor of dtype int32 which PyTorch does not like. For future reference, the error:

Expected tensor for argument #1 'xxx' to have scalar type Long; but got torch.IntTensor instead

is usually a clear indicator of this behavior happening.

The PR fixes the QA pipeline by casting the tensors to long if they have the int type.

[LysandreJik]

LGTM, thanks @sgugger!