Fix three bugs in vocoder loss computation

[Mr-Neutr0n]

Summary

Fixes three bugs in TTS/vocoder/layers/losses.py that cause incorrect loss values during vocoder training:

• Boolean logic error in GeneratorLoss.forward(): The condition if self.use_hinge_gan_loss and not scores_fake is not None evaluates as (not scores_fake) is not None due to Python operator precedence, which is always True. This caused the hinge loss block to execute even when scores_fake was None, leading to a runtime error. Fixed to and scores_fake is not None. Applied the same fix to the feat_match_loss guard which had an analogous issue (and not feats_fake: is falsy for empty lists).

• Variable shadowing in _apply_D_loss(): The loop unpacking total_loss, real_loss, fake_loss = loss_func(...) shadows the accumulator variables real_loss and fake_loss initialized above. As a result, real_loss += real_loss just doubles the current iteration's value instead of accumulating across discriminator scales. Renamed the unpacked variables to cur_real_loss and cur_fake_loss.

• Feature loss normalization in MelganFeatureLoss.forward(): loss_feats /= len(fake_feats) + len(real_feats) divides by 2N instead of N, since both feature lists have the same length. The loss is summed over N pairs, so the denominator should be N. Fixed to loss_feats /= len(fake_feats).

Test plan

• Verify hinge GAN loss is correctly applied when use_hinge_gan_loss=True and scores_fake is provided
• Verify feature matching loss is correctly applied when use_feat_match_loss=True and feats_fake is provided
• Verify _apply_D_loss correctly accumulates real_loss and fake_loss across multiple discriminator scales
• Verify MelganFeatureLoss returns correctly normalized loss values