Fix a noise shape of StyleMelGANGenerator to export ONNX model

[c-bata]

Hi! Thank you for creating this awesome project. In this PR, I suggest modifying StyleMelGANGenerator to export the ONNX model.

Problem details

I bumped the following issue when I tried to export StyleMelGANGenerator.

/path/to/venv/lib/python3.8/site-packages/parallel_wavegan/models/style_melgan.py:222: TracerWarning: Converting a tensor to a Python float might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  math.ceil(c.size(2) / self.noise_upsample_factor),

This warning shows that math.ceil() operation cannot be traced by torch.onnx.export() function. So that the shape of noise will be treated as constant values.

After applying this patch, we can export ONNX model.

[c-bata]

Note that both of c.size(2) and self.noise_upsample_factor are intergers. So that we can replace math.ceil() operation like this. Even though we may use torch.ceil() function, it takes and returns a tensor, not a scalar value.

>>> import math
>>> noise_upsample_factor = 5
>>> for c_size in range(15):
...     before = math.ceil(c_size / noise_upsample_factor)
...     after = (c_size-1) // noise_upsample_factor + 1
...     print(f"before: {before}, after: {after}")
...
before: 0, after: 0
before: 1, after: 1
before: 1, after: 1
before: 1, after: 1
before: 1, after: 1
before: 1, after: 1
before: 2, after: 2
before: 2, after: 2
before: 2, after: 2
before: 2, after: 2
before: 2, after: 2
before: 3, after: 3
before: 3, after: 3
before: 3, after: 3
before: 3, after: 3

[kan-bayashi]

Thank you for your modification! It looks fine.
I slightly modified to pass the CI.