[c10] Fix the hanlding for Caffe2 ops which return tensor list (#36841)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36841
right now, all c2 ops's output will be unwrapped blindly. This is not correct, if we have a single tensor list returned.
Test Plan: buck test mode/dev-nosan mode/no-gpu //caffe2/caffe2/fb/python/operator_test:torch_integration_test
Reviewed By: alyssawangqq
Differential Revision: D21100463
fbshipit-source-id: 9f22f3ddf029e7da9d98008d68820bf7f8239d4f