feat: Refactor training with SOLID principles and add optimizations

feat: Add multi-GPU support and lmdeploy for faster TTS

feat: Add H200/Hopper support and VRAM-adaptive batch sizing

fix: Update dependencies and tokenizer handling for dataset generation

perf: Increase TTS_BATCH_SIZE to 200 for better throughput

feat: Pre-compute text_tokens in dataset generation

feat: Async dataset pipeline with adaptive memory management