[TTS] Audio codec fixes

[rlangman]

What does this PR do ?

A few minor fixes for the audio codec training recipe.

Most important, moving the zero_gradient() call back before the loss calculations. Having it after optim_disc.step() results in it saving discriminator gradients during the generator self.manual_backward(loss_gen_all) calculation which stops the discriminator from learning (adversarial and feature losses quickly become 0).

Collection: [TTS]

Changelog

• Zero gradient before discriminator loss calculation
• Change default mel loss scale from 1.0 to 5.0 as it significantly improves ViSQOL and makes training more consistent.
• Fix name of validation time domain loss

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation