Fix WebSocket streaming - ParlerTTSStreamer yields numpy arrays not tensors

Add WebSocket endpoint for true real-time bidirectional streaming

Implement official ParlerTTSStreamer for true real-time audio streaming

Revert to sentence-by-sentence processing with fp16 - preserves word quality while maintaining speed

Optimize TTS for speed - add fp16, batch processing, remove sentence splitting

Add HF_TOKEN authentication for gated model access

Optimize CUDA device handling - auto-detect GPU and remove unnecessary device transfers

Initial commit: Parler TTS FastAPI with Docker