[caffe2] improve core.Net cloning/init performance (24x for large models!) (#47475)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/47475
This improves the core.Net cloning/init performance by quite a bit. It makes set_input_record run in linear time instead of O(n) by checking the external_input map instead of regenerating the external inputs each time and then iterating over it.
Test Plan: unit tests + canary runs
Reviewed By: dzhulgakov
Differential Revision: D24765346
fbshipit-source-id: 92d9f6dec158512bd50513b78675174686f0f411
