Change Grad to return zero tensors instead of nan tensors when applied to an empty list of outputs

Author: ValerianRey

src/torchjd/autojac/_transform/grad.py

@@ -37,12 +37,7 @@ def _differentiate(self, grad_outputs: Sequence[Tensor]) -> tuple[Tensor, ...]:
             return tuple()
 
         if len(outputs) == 0:
-            return tuple(
-                [
-                    torch.empty(input.shape, device=input.device, dtype=input.dtype)
-                    for input in inputs
-                ]
-            )
+            return tuple([torch.zeros_like(input) for input in inputs])
 
         optional_grads = torch.autograd.grad(
             outputs,