Increase sampling rate to 24000 for training

[Darwinkel]

Make it the same as XTTSv2 output - but requires more compute power and results in misalignment between original training checkpoint. Initial results not promising.