Trainer refactor, Vits multi-GPU fixes, transformer upgrades

[eginhard]

• Update to coqui-tts-trainer 0.4.0, which allows sharing more code between models or unifying implementations.
• Use soundfile to check audio length, which is significantly faster than torchcodec.
• Fix multi-GPU support for custom samplers in Vits training. Fixes Do multigpu training with weight sampler work with vits in this repo. #103, How to properly run on multiple GPUs coqui-ai/TTS#2542, Multi-GPU training YourTTS model fail coqui-ai/TTS#2584, [Bug] Error Dataloader when using distributed training vtck vits coqui-ai/TTS#3596
• Move speaker encoder to correct device. Fixes [Bug] gpu flag is marked deprecated but still affects device usage in YourTTS coqui-ai/TTS#4398
• Add support for transformers 5.1, fixes [Bug] coqui-ai-tts is not compatible with transformers>=5.1 #558, [Bug] ImportError: cannot import name 'isin_mps_friendly' from 'transformers.pytorch_utils' coqui-ai/TTS#4406
• fix(bark): check if voice_dir exists before reading it
• fix(vits): convert tensors to float32 first