feat: batched graph construction that also support GPU

[yanghan234]

Summary

This PR implements a batched graph generation pipeline that supports GPU acceleration, alongside general code cleanup and refactoring.

Key Changes

• New Graph Converter: Introduced BatchGraphConverter in converter.py. This new converter leverages batch processing and can utilize CUDA for faster graph construction.
• Three-Body Acceleration: Added threebody_indices_torch.py to compute three-body indices efficiently on the GPU, replacing the CPU-bound iteration.
• Refactored build_dataloader
  • Updated build_dataloader in build.py to support the new pipeline.
  • Added a batch_converter flag (default: True) to easily switch between the new optimized batch converter and the legacy single-item converter.
• Dependencies: Updated pyproject.toml to include torch-scatter and refine torch version.
• Add a script scripts/benchmark_graph_converter.py to benchmark and compare the speed of the new graph implementation

[yanghan234]

By running the benchmark script on the following structures: test_structures.xyz.zip, you should be able to see the following results:

----------------------------------------
Energy Error Summary (batch_converter=True vs False):
Number of structures: 1000
MAE:           7.675171e-06 eV
RMSE:          3.186059e-05 eV
Max Error:     4.882812e-04 eV
MAE/atom:      2.088059e-04 meV/atom
RMSE/atom:     4.678199e-04 meV/atom
Max Error/atom: 4.577637e-03 meV/atom
----------------------------------------
Timing Summary:
Graph conversion (ref):  1.8792s
Graph conversion (test): 0.9001s
Speedup (graph only):    2.09x
Total time (ref):        3.0376s
Total time (test):       1.3849s
Speedup (total):         2.19x
----------------------------------------

Also, the difference of the energy prediction and the time speedup are shown as plots: