Support Intel XPU

[richiejp]

What this does

While testinging on Intel (XPU) I found that:

• 2_evaluate_pretrained_policy.py failed with an error claiming tensors were on CPU as well as XPU. Moving the policy to the specified device (xpu) fixes the issue and matches the behaviour in the training script.
• train.py failed to run just because some utility functions failed to recognize XPU as a valid device type, so I added it.

Also this adds a Dockerfile file for XPU based on Intel's PyTorch Dockerfile.

How it was tested

I ran the policy training and evaluation for pusht on an Intel Arc a770 using the provided Dockerfile.

How to checkout & try? (for the reviewer)

Run the examples/2_evaluate_pretrained_policy.py script or the training script.

To test on Intel the following Docker commands could be used:

docker build -t lerobot:xpu -f docker/lerobot-xpu/Dockerfile .
docker run -it --rm \
      --device /dev/dri \
      -v /dev/dri/by-path:/dev/dri/by-path \
      --ipc=host \
      -v ~/.cache/huggingface:/root/.cache/huggingface \
      -v ~/python/lerobot/outputs:/lerobot/outputs \
      lerobot:xpu python lerobot/scripts/train.py \
      --dataset.repo_id=lerobot/pusht \
      --policy.type=diffusion \
      --policy.device=xpu \
      --env.type=pusht \
      --log_freq=100 \
      --save_freq=1000 \
      --steps=10000

[richiejp]

While trying to run SmolVLA I found XPU does not support f64, so allowed it to fallback to f32 the same as mps.