[dist_optim] update the doc of DistributedOptimizer

updating the doc of DistributedOptimizer to include TorchScript enablement information

ghstack-source-id: 2fd88164f11550a21d8bf3e4f119c76e50641204
Pull Request resolved: #51314

torch/distributed/optim/optimizer.py

@@ -150,6 +150,13 @@ class DistributedOptimizer:
     to the latest forward pass executed on a given worker. Also, there is no
     guaranteed ordering across workers.
 
+    `DistributedOptimizer` creates the local optimizer with TorchScript enabled
+    by default, so that optimizer updates are not blocked by the Python Global
+    Interpreter Lock (GIL) during multithreaded training (e.g. Distributed Model
+    Parallel). This feature is currently in beta stage, enabled for optimizers
+    including `Adagrad`, `Adam`, `SGD`, `RMSprop`, `AdamW` and `Adadelta`. We
+    are increasing the coverage to all optimizers in future releases.
+
     Args:
         optimizer_class (optim.Optimizer): the class of optimizer to
             instantiate on each worker.