docs(mge/python): add doc for zero_grad

GitOrigin-RevId: bb5154c0e7280bfb3162d5a34c508df4c0be2847

python_module/megengine/functional/utils.py

@@ -69,4 +69,12 @@ def accuracy(logits: Tensor, target: Tensor, topk: Union[int, Iterable[int]] = 1
 
 @wrap_io_tensor
 def zero_grad(inp: Tensor) -> Tensor:
+    r"""
+    Returns a tensor which is treated as constant during backward gradient calcuation,
+    i.e. its gradient is zero.
+
+    :param inp: Input tensor.
+
+    See implementation of :func:`~.softmax` for example.
+    """
     return mgb.opr.zero_grad(inp)