fix: move tensors to correct device in batch_to_dict

[yanghan234]

Summary

Fixes #113 — batch_to_dict accepted a device parameter but never used it, leaving all tensors on their original device.

Changes

• Add .to(device) to every tensor extracted from graph_batch in batch_to_dict()
• Create num_graphs tensor with device=device
• Pass self.device at all 5 internal call sites so the correct runtime device (cpu/mps/cuda) is always used

Tests

Added tests/forcefield/test_batch_to_dict.py with 9 test cases covering:

• Tensor placement on all available devices (CPU, MPS, CUDA)
• Cross-device movement
• Value preservation after device transfer
• num_graphs scalar-to-tensor conversion on correct device
• Unsupported model type error handling

Also added tests/conftest.py with a shared device fixture that auto-detects available devices.