rc4 qat finetune data loading

Summary:
With this diff, we are able to reproce the data loading based on the final rc4 model config

## Changes
1. Support category frequency as per instance weights instead of dataset level weights
2. Update rc4 qat finetune model config
3. Use naive sync BN in rc4 backbone

Reviewed By: stephenyan1231

Differential Revision: D56121350

fbshipit-source-id: 554446995ec9435ce439d7e454fe6ccf84a30934

detectron2/data/build.py

@@ -484,7 +484,7 @@ def _train_loader_from_config(cfg, mapper=None, *, dataset=None, sampler=None):
                 repeat_factors = RepeatFactorTrainingSampler.repeat_factors_from_category_frequency(
                     dataset, cfg.DATALOADER.REPEAT_THRESHOLD, sqrt=cfg.DATALOADER.REPEAT_SQRT
                 )
-                sampler = RepeatFactorTrainingSampler(repeat_factors)
+                sampler = RepeatFactorTrainingSampler(repeat_factors, seed=cfg.SEED)
             elif sampler_name == "RandomSubsetTrainingSampler":
                 sampler = RandomSubsetTrainingSampler(
                     len(dataset), cfg.DATALOADER.RANDOM_SUBSET_RATIO