Add CUequivariance backend

[CompRhys]

How hard would it be to add a CUequivariance back end to SevenNet?

[YutackPark]

It was harder than I firstly thought. Now available on cu_euqi branch.

Nearly x3 times faster in training, when the channel is consistent across the L value. If # of channels is different for each L value, like SevenNet-0, I found it becomes slightly slower.
Slightly (~50%?) faster for MD even with an inconsistent number of channels.

Here's the limitations and usage training.

Issues

1. SevenNet-0 is not available. Due to this issue:
   [Bug] torch.autograd.grad never ends, if ChannelWiseTensorProduct is used with specific configuration NVIDIA/cuEquivariance#33
   Currently, sevennet automatically fallback to the original e3nn implementation if the condition is met. Models with higher Lmax (>2) will work.

2. LAMMPS is not available. Due to this issue:
   Questions about TorchScript support NVIDIA/cuEquivariance#30
   It is a more unclear than the previous limitation. PyTorch itself may deprecated the torchscript than we have to re-implement LAMMPS part when newer replacement, torch.export came out. I'm not sure whether they gonna support torchscript for cuEquivariance.

Usage:

Installation:

git clone https://github.com/MDIL-SNU/SevenNet.git
cd SevenNet
git checkout cu_equi
pip install .

pip install cuequivariance-torch
pip install cuequivariance-ops-torch-cu12  # choose between cu12 or cu11 based on CUDA version

Training:

sevenn input.yaml --enable_cueq

Calculator:

from sevenn.sevennet_calculator import SevenNetCalculator
calc = SevenNetCalculator(PATH_TO_CHECKPOINT, enable_cueq=True)

The interface is intentionally made similar to that of MACE.

[YutackPark]

#169
Pre-release available. Now it is actually useful as recent updates of cuEquivariance boosted 7net-0.