Small clarification of torch.cuda.amp multi-model example (#41203)

Summary:
some people have been confused by `retain_graph` in the snippet, they thought it was an additional requirement imposed by amp.

Pull Request resolved: pytorch/pytorch#41203

Differential Revision: D22463700

Pulled By: ngimel

fbshipit-source-id: e6fc8871be2bf0ecc1794b1c6f5ea99af922bf7e

docs/source/notes/amp_examples.rst

@@ -254,6 +254,8 @@ after all optimizers used this iteration have been stepped::
                 loss0 = loss_fn(2 * output0 + 3 * output1, target)
                 loss1 = loss_fn(3 * output0 - 5 * output1, target)
 
+            # (retain_graph here is unrelated to amp, it's present because in this
+            # example, both backward() calls share some sections of graph.)
             scaler.scale(loss0).backward(retain_graph=True)
             scaler.scale(loss1).backward()