fix the bug where convert bn to sync_bn in the model for evaluation during the training phase will change the training model graph

[FlyingQianMM]

The origin implementation converts bn to sync_bn in the model for evaluation during the multi-gpu training phase, which is not necessarily and changes the training model graph, causes the gradient of multi-gpu cannot be gathered during later training backward:

so we remove the convertion.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}