Enhance decorator _use_grad_for_differentiable (pytorch#103567)

Aim: enhance decorator _use_grad_for_differentiable so that functions (methods) decorated by it keep their docstrings and signatures unchanged.

Fixes pytorch#103566

Pull Request resolved: pytorch#103567
Approved by: https://github.com/janeyx99

torch/optim/optimizer.py

@@ -38,6 +38,7 @@ def _use_grad(self, *args, **kwargs):
         finally:
             torch.set_grad_enabled(prev_grad)
         return ret
+    functools.update_wrapper(_use_grad, func)
     return _use_grad
 
 def _get_value(x):