🐛 Bug

repro:

1. have a model with nn.BatchNorm2d which is not intended to be fused
2. run torch.quantization.prepare_qat on the model to prepare for QAT
3. run torch.nn.SyncBatchNorm.convert_sync_batchnorm on the model to prepare for DDP
4. run torch.quantization.convert_qat to sub in the quantized modules

Currently step 3 does not carry over the qconfig, so the BNs never get quantized in step 4. We should make sure that
a. qconfig survives the swaps.
b. either convert works with SyncBN directly, or there is a utility to go from SyncBN back to BN (tracked in #41081)

cc @jerryzh168 @jianyuh @dzhulgakov @raghuramank100 @jamesr66a @vkuzo