Standardize order of tensor dimensions

[Waino]

OpenNMT uses the order seq batch dim, while x-transformers uses batch seq dim. Currently, we transpose tensors to limit the x-transformers order to the parts where it is strictly necessary, while retaining the OpenNMT order everywhere else.

To simplify the code and avoid unnecessary transposes, we should use batch seq dim everywhere, starting from the data loader all the way through to the translation decoding.