Preserve autograd history in torch backend

scoringrules/backend/registry.py

@@ -59,11 +59,9 @@ def __getitem__(self, __key: str) -> ArrayBackend:
         """Get a backend from the registry."""
         try:
             return super().__getitem__(__key)
-        except KeyError as err:
-            raise BackendNotRegistered(
-                f"The backend '{__key}' is not registered. "
-                f"You can register it with scoringrules.register_backend('{__key}')"
-            ) from err
+        except KeyError:
+            self.register_backend(__key)
+            return super().__getitem__(__key)
 
     def set_active(self, backend: str):
         self._active = backend

scoringrules/backend/torch.py

@@ -32,7 +32,8 @@ def asarray(
         *,
         dtype: Dtype | None = None,
     ) -> "Tensor":
-        return torch.asarray(obj, dtype=dtype)
+        # torch.asarray(obj) would cancel gradients!
+        return torch.as_tensor(obj, dtype=dtype)
 
     def broadcast_arrays(self, *arrays: "Tensor") -> tuple["Tensor", ...]:
         return torch.broadcast_tensors(*arrays)