fix: Solve SmolVLA meta tensor error by removing device_map

[Highsky7]

Fixes meta tensor copy issue in factory.py:418.

What this does

This PR fixes a NotImplementedError: Cannot copy out of meta tensor that occurs when training the SmolVLA policy with load_vlm_weights=True.

The error happened because device_map=device in SmolVLMWithExpertModel caused accelerate to place some parameters on the meta device. LeRobot's factory later attempts to move the entire policy using .to(), which fails for meta tensors.

Changes:

• Remove device_map parameter from VLM model loading to prevent automatic meta device placement.
• Change torch_dtype from string "bfloat16" to torch.bfloat16 object.
• Add explicit .to(device) calls after initialization to ensure the model and lm_expert are correctly placed on the target device.

This resolves NotImplementedError when training SmolVLA policy.

How it was tested

Tested locally with the LibERO dataset (libero_10 task).

• Verified that training starts successfully without the meta tensor error.
• Verified that the model loads to CUDA correctly.
• Verified that loss computation works as expected for 10 steps.

How to checkout & try? (for the reviewer)

python3 src/lerobot/scripts/lerobot_train.py \
  --policy.type=smolvla \
  --policy.load_vlm_weights=true \
  --policy.device=cuda \
  --dataset.repo_id=HuggingFaceVLA/libero \
  --env.type=libero \
  --env.task=libero_10 \
  --output_dir=./outputs/train/test \
  --steps=10