fix recording bug in after_train function

[Coolshanlan]

When I evaluate model in the after_train function and record results, the result will not be recorded in Tensorboard and metrics.json

This is because detectron2.utils.events.TensorboardXWriter and JSONWriter will check the trainer.iter, and each iter will only record once.
If you do not set trainer.iter+1 before executing after_train, the execution result will not be recorded, because the current trainer.iter has been used in the last step

So I fixed this issue in train_loop function