Avoid keeping old histograms in the histogram observer to fix the OOM issue (#29768)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29768
The previous histogram observer saves all histograms for new data and merge the histograms in the end. It could cause OOM issue when we want to collect histograms on large amount of data. In this diff, we assume running the histogram observer with a single thread and remap the histogram after seeing new data.
Test Plan:
```
buck test mode/opt caffe2/caffe2/quantization/server:dynamic_histogram_test
```
```
buck run mode/opt caffe2/caffe2/fb/fbgemm/numerical_debugger/workflows:int8_static_quantization_exporter -- --model-dir /mnt/public/summerdeng/ads/ --model-name downsized_ins_97293388_0.predictor --run --iter 10 --dataset-path /mnt/public/summerdeng/ads/ctr_instagram_story_int8/dataset/train/dataset_115764229_10 --hive-path="hive://ad_delivery/ig_ad_prefiltered_training_data_orc_injected/ds=2019-09-09/pipeline=ctr_instagram_story_click_only_model_opt_out_df" --collect-histogram --activation-histogram-file=/mnt/public/summerdeng/ads/ctr_instagram_story_int8/activation_histograms/dummy_debug_OOM.txt
```
Reviewed By: jspark1105
Differential Revision: D18458764
fbshipit-source-id: c0e36fffe9bf021efd17d8494deef43727333da2
