Fix automatic CUDA graphing not working when requiring backwards (#120)

* Pass posTensor as input argument to energyTensor.backwards.
This instructs torch to compute gradients only with respect to positions.

* Add comments

Author: RaulPPelaez

platforms/cuda/src/CudaTorchKernels.cpp

@@ -188,7 +188,11 @@ static void executeGraph(bool outputsForces, bool includeForces, torch::jit::scr
         energyTensor = module.forward(inputs).toTensor();
         // Compute force by backpropagating the PyTorch model
         if (includeForces) {
-            energyTensor.backward();
+            // CUDA graph capture sometimes fails if backwards is not explicitly requested w.r.t positions
+	    // See https://github.com/openmm/openmm-torch/pull/120/
+	    auto none = torch::Tensor();
+	    energyTensor.backward(none, false, false, posTensor);
+	    // This is minus the forces, we change the sign later on
             forceTensor = posTensor.grad().clone();
             // Zero the gradient to avoid accumulating it
             posTensor.grad().zero_();