Various bug fixes

[lars76]

I was finetuning the models on my own datasets and it failed due to a couple of bugs. This pull request fixes these issues:

• only divide by MAX_WAV_VALUE if the wav file is saved as int16 (and not float).
• always try to pad to the length of frames_per_seg or self.segment_size. It can happen that audio.size(1) < frames_per_seg * self.hop_size
• first comes tqdm, then enumerate, otherwise tqdm is not able to get the length attribute
• Make sure that y_mel, y_g_hat_mel, y_g_hat, y_g_hat_mel are the same length. I think the issue is that sometimes my wav files have uneven length.

[L0SG]

Thank you for your contribution. Looks correct when fine-tuning using a model-generated mel (like TTS). Also based on your feedback, current implementation of load_wav() is indeed not robust to load heterogeneous waveform formats. I'll merge it including my additional commits to provide better robustness loading the waveforms soon.

[L0SG]

Merged with my additional implementation to support loading waveform with various file formats, including on-the-fly resampling.

Note regarding fine-tuning (using pre-computed mel): It is recommended to prepare the waveform pair with length of integer multiple of the target hop length to ensure precise alignment.

If you find any additional bugs of the new implementation, please let me know. Thanks!