FIX potential memory error with torch.ones in tts_utils.py

[HenraL]

Passed the torch.ones to torch long instead of torch.32 so that they are in the same size as the tokens that are passed to the model.