Call contiguous() after permute() in the pytorch model

[hqucms]

This avoids an error in some versions (e.g., 1.4) of PyTorch, and also brings sizable speed improvement.