[TensorPipe] Always complete futures from thread pool (#38930)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/38930
Any time we mark a future as complete or set an error on it we call its callbacks, which could be arbitrary user functions and could thus be slow or blocking. The safest behavior is to always defer to the loop.
ghstack-source-id: 104760682
Test Plan: None... :(
Differential Revision: D21703017
fbshipit-source-id: ad2bdc6be25844628ae6f318ef98b496f3d93ffd