Support CUDA Graphs

[raimis]

Inspired by the significant performance increase with CUDA Graphs (torchmd/torchmd-net#48 (comment)), this PR implemented the support of CUDA Graphs to accelerate PyTorch models (https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/).

• Separate the graphable operations
• Make the input and output tensors persistent
• Capture the graph
• Expose an option to enable the graph
• Serialize the options
• Conditionally build for the old PyTorch version without the support of CUDA Graphs
• Add tests
• Update documentation

[raimis]

@peastman I need your opinion.

We need to extend the API for platform-specific options, e.g. the usage of CUDA Graphs.

My proposal is to add a simple dictionary-style API:

TorchForce.getPropertyValue(name: str) -> str
TorchForce.setPropertyValue(name: str, value: str)

For convenience, the constructor can accept a dictionary directly:

TorchForce(filename: str, options: Optional[Dict[str, str]] = None)

The names and values are checked when a platform-specific CalcTorchForceKernel is initialized.

[peastman]

That sounds reasonable.

[raimis]

Deprecated by #103