Prefer reshape to view in computations

[odusseys]

Thanks for your great work!
A slight issue in several functions of the library is that they use .view() to reshape inputs, which fails on non-contiguous tensors. For instance, the sinkhorn divergence scaling_parameters function fails if two noncontiguous tensors are passed to the loss function, which is unexpected. Using reshape instead would allow bypassing the need to explicitly make tensors contiguous before passing to SamplesLoss, while still creating views when possible and not sacrificing performance.