Split out custom ops lib from executorch_generated_lib

Summary:
Currently `executorch_generated_lib` macro generates 2 c++ libs:
1. C++ lib for all ops (ATen + custom ops) registration into Executorch runtime.
2. C++ lib for custom op registration into PyTorch.

2 is used by EXIR (AOT) and 1 is used by runtime. This is causing issues in OSS because 2 is required by EXIR but also needs to depend on pytorch C++ library. We want to split it out and figure out a good way to support custom ops registration in OSS.

This diff is a pure refactoring, where we extract out 2 into its own macro. In next diff I'm going to create the targets for `exir_custom_ops_aot_lib` separately.

Reviewed By: cccclai

Differential Revision: D47976256

fbshipit-source-id: db0031a96817600b2e545fefcfe3c269bc34cf54