[minor] Update wavenet.py

[p0p4k]

1. (based on https://github.com/jaywalnut310/vits)
2. In the earlier PR Update wavenet.py #1796 , I made a mistake about understanding the WaveNet architecture.
3. If we follow the current implementation, hidden_channels will always have to be equal to in_channels. For a general use of this model, I had changed the res_skip_layers to output 2 x in_channels at every except the last layer and in_channels at the last layer.
4. This change will not break existing GlowTTS and VITS models as they use num_hidden_channels == num_in_channels for their respective implementations.
5. (minor|experimental) Having said that, there could be a use of dropout after the fused tanh sigmoid activation layer, which I have added as a comment for experimentation with a warning that it will break BC if used.