Add Intel XPU (Arc GPU) inference support

[munikera]

Summary

Adds inference support for Intel Arc GPUs via the PyTorch XPU backend. Tested on two Intel Arc B70 GPUs (Ubuntu 22.04, xe kernel driver). Achieves ~130 ms/step on pi05_libero.

Changes

• policy_config.py: auto-detect XPU device in create_trained_policy (falls back to xpu when CUDA is unavailable)
• pyproject.toml: add xpu dependency group with torch==2.11.0+xpu, torchvision==0.26.0+xpu, triton-xpu==3.7.0 from the PyTorch XPU index; loosen torch==2.7.1 -> torch>=2.7.1 to allow the XPU wheel to satisfy the base constraint
• docs/xpu.md: end-to-end setup guide covering driver install, dependency setup, checkpoint download/conversion, smoke test, and LIBERO benchmark eval

Note on torch pin

The base torch==2.7.1 pin is loosened to torch>=2.7.1 because uv must satisfy both the base constraint and the XPU extra (torch==2.11.0+xpu) simultaneously. This does
not affect non-XPU installs - uv sync without --group xpu still resolves torch==2.7.1.

Verification

# Install XPU dependencies
uv sync --group xpu

# Verify XPU is detected
uv run python -c "import torch; print(torch.xpu.is_available())"

# Smoke test (requires policy server running)
uv run scripts/serve_policy.py --port 8000 policy:checkpoint \
    --policy.config pi05_libero \
    --policy.dir checkpoints/pi05_libero_pytorch

uv run examples/simple_client/main.py --env LIBERO --num-steps 5

See docs/xpu.md for full setup instructions.