Add missing .float() for bf16 inference for FSQ models

[NilanEkanayake]

Simple change that allows BF 16 inference like so:

model.to(torch.bfloat16)
with torch.no_grad(), torch.autocast(device_type='cuda:0', dtype=torch.bfloat16):
    _, xrec, _ = model(input)

...although loss calculation needn't be run during inference anyways.

[deeptimhe]

Hi, thanks for the contribution! We are working on testing the loss for bfloat16 inference.