fixed noise of unvoiced segment (fix #30, #17)

• Significant improvement in unvoiced segments (see #30, #17)
• Match details of official implementation to improve quality
• New pre-trained model
• New audio samples at http://swpark.me/melgan/.

MelGAN v0.2-alpha

Fixed wrong implementation of Generator (see #22)

PyTorch Hub release test

v0.1-alpha

remove weight_norm, add torch hub