Fix vits low-precision dtype

[jiqing-feng]

This PR fixed vits model dtype when torch_dtype=torch.float16.

To reproduce the error:

import torch
from transformers import pipeline

pipe = pipeline("text-to-speech", model="facebook/mms-tts-eng", torch_dtype=torch.float16)
output = pipe("Hello, my dog is cooler than you!")
print(output)

Traceback:

......
  File "/home/jiqing/transformers/src/transformers/models/vits/modeling_vits.py", line 916, in forward
    query_states = self.q_proj(hidden_states) * self.scaling
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1747, in _call_impl
    return forward_call(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/linear.py", line 125, in forward
    return F.linear(input, self.weight, self.bias)
RuntimeError: mat1 and mat2 must have the same dtype, but got Float and Half