Change device but not dtype in UnitY2AlignmentModel

[d-dawg78]

When embs_text.size(1) or embs_unit.size(1) is large and the model's type is torch.float16, there are half-precision issues. This is easily reproducible locally:

>>> import torch
>>> torch.tensor([2131]).to(dtype=torch.float16).int()
tensor([2132], dtype=torch.int32)

This causes downstream issues, especially in this line. We got the following error in production:

score = score.masked_fill(padding_mask, -np.inf)
RuntimeError: The size of tensor a (2136) must match the size of tensor b (2137) at non-singleton dimension 2

The following fix aims to make sure that there are no unnecessary dtype conversions in the code that could cause these size mismatch issues. We do however make sure that the device is changed if need be.