Demo: Fix hardcoded CUDA references to support Apple Silicon MPS and fix input_sample_rate

[kainoj]

Summary

• Auto-detect available device (cuda → mps → cpu) in the demo instead of assuming CUDA
• Pass the detected device through to all from_pretrained calls and the Mimi warmup loop
• Work around safetensors not supporting MPS as a direct load target by falling back to CPU load when
  the device is neither cpu nor cuda
• Replace hardcoded .cuda() on the audio detokenizer with .to(self.device)
• Fix input_sample_rate from 24_000 to 48_000 to match Chrome's native WebRTC rate — avoids a lossy
  48→24→16kHz double resample before the conformer encoder (which expects 16kHz).
  • This subjectively improved the experience. Before the change the model was replying "sorry i couldn't hear you clearly, could you repeat" etc

Test plan

• Run liquid-audio-demo on Apple Silicon Mac — should load and run on MPS without errors