Free original weight after prepacking in XNNPACK based op (#46541)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/46541
When weights are prepacked XNNPACK packs the weights in a separate memory. After that original weights are not needed for inference. Having those weights lying around increase memory footprint, so we would like to remove the original weights once prepacking is done.
Test Plan: buck test //caffe2/aten:mobile_memory_cleanup
Reviewed By: kimishpatel
Differential Revision: D24280928
fbshipit-source-id: 90ffc53b1eabdc545a3ccffcd17fa3137d500cbb
