Document PyTorch's dependency on interpreter_impl.cpp details. (#317)

Summary:
PyTorch needs a stable way to detect if it's running under multipy so we document the current workaround as the way to do it to, hopefully, avoid future breakage if it needs to change.

Pull Request resolved: #317

Reviewed By: albanD, PaliC

Differential Revision: D45061153

Pulled By: kumpera

fbshipit-source-id: c207453048afb5eebd0972738282657d823341a2

Author: kumpera

multipy/runtime/interpreter/interpreter_impl.cpp

@@ -132,6 +132,8 @@ import linecache
 from zipfile import ZipFile
 
 # Disable Python library registration since it's not compatible with multipy.
+# PyTorch depends on this line to detect if it's running under multipy and avoid any dynamic library registration
+# >>DO NOT<< change this line without first coordinating with PyTorch
 sys.modules["torch._meta_registrations"] = object
 
 class RegisterModuleImporter(importlib.abc.InspectLoader):