Enable XPU inference for InternVLA-M1 (backend-safe)

[padatta]

Summary

Enables InternVLA-M1 inference on Intel XPU.

Changes

• InternVLA/model/modules/vlm/QWen2_5.py
  • Add _get_autocast_device() helper (reads the model's device).
  • Load Qwen2.5-VL with backend-safe attention + placement: flash_attention_2 and device_map="cuda" on CUDA; sdpa and device_map="xpu"/"cpu" otherwise.
  • Replace hard-coded torch.autocast("cuda", ...) with the detected device (2 sites).
• InternVLA/model/framework/M1.py
  • Replace hard-coded torch.autocast("cuda", ...) with the detected device (4 sites).
  • chat_with_M1(): device defaults to the model's own device instead of a fixed string.
  • main device selection prefers XPU, then CUDA, then CPU.

Compatibility

• CUDA: unchanged
• XPU: newly supported

Testing

• python -m py_compile on changed files.
• Inference smoke test on Intel XPU.