Improve error message of pipeline (#3260)

Summary:
Pull Request resolved: #3260

For https://fb.workplace.com/groups/260102303573409/permalink/737763939140574/

Clarifying that user should detach pipeline before no_grad or model forward call

Reviewed By: jd7-tr

Differential Revision: D79728704

fbshipit-source-id: ebcb418e56ee11d7b41a1aa730b49512a2507d43

Author: kausv

torchrec/distributed/train_pipeline/runtime_forwards.py

@@ -121,7 +121,7 @@ def __call__(
     ]:
         assert (
             self._name in self._context.embedding_a2a_requests
-        ), "Invalid EmbeddingPipelinedForward usage, please do not directly call model.forward()"
+        ), "Invalid EmbeddingPipelinedForward usage, please call pipeline.detach() before torch.no_grad() and/or  model.forward()"
 
         ctx = self._context.module_contexts.pop(self._name)
         cur_stream = torch.get_device_module(self._device).current_stream()