Remove unpacked int8 blob after constructing the packed blob to save memory (#37973)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37973
Fix the unexpected memory usage issue in model QRT for the OC model.
Test Plan:
```
buck test mode/opt caffe2/caffe2/quantization/server:fully_connected_dnnlowp_op_test
```
```
buck test mode/opt caffe2/caffe2/fb/fbgemm:int8_serializer_test
```
Reviewed By: hx89
Differential Revision: D21422257
fbshipit-source-id: cc586123b8bfe41c85c6f2f7e493954845ad18a2
