[StyleTTS2] Fix training checkpointing

[roedoejet]

Previously, training would log every epoch. Now, it mimics FS2 - that is, log the last epoch as well as the top 5 best ckpts in terms of Validation Mel Spectrogram loss. It will also fix the audio-logging issue where only one audio file was being logged to tensorboard. I also added text logging.

To test, try running everyvoice train and inspect the tensorboard logs.

EveryVoiceTTS/StyleTTS2#5

[semanticdiff-com]

Review changes with  

Changed Files

File	Status
everyvoice/model/e2e/StyleTTS2_lightning	0% smaller

[github-actions]

CLI load time: 0:00.37
Pull Request HEAD: 11d27afebebf0d73a6e2593e5a348300ac354253

Imports that take more than 0.1 s:
import time: self [us] | cumulative | imported package
import time:      2043 |     191879 | everyvoice.cli

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.81%. Comparing base (2e00485) to head (11d27af).

Additional details and impacted files

@@                      Coverage Diff                       @@
##           dev.ap/styletts2-improvements     #805   +/-   ##
==============================================================
  Coverage                          84.81%   84.81%           
==============================================================
  Files                                 46       46           
  Lines                               4189     4189           
  Branches                             626      626           
==============================================================
  Hits                                3553     3553           
  Misses                               499      499           
  Partials                             137      137           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.