CUDA support

[kalihale]

When using a graphics card with CUDA support, there is an unhandled tensor error where the tensor must be moved to the CPU before converting to a numpy array.

substrafl/substrafl/algorithms/pytorch/torch_newton_raphson_algo.py

Line 140 in 95ecc62

final_gradients = [torch.zeros_like(p).numpy() for p in self._model.parameters()]