[Bug-fix] Set batch size min to world size

[KSGulin]

There's currently a bug with scaling down batch size for QAT, as batch size must be a multiple of the world size but it can be set to 1 when scaling down. This fix updates the batch size floor to be equal to the world size.

Test plan
Tested locally with command

python -m torch.distributed.run --no_python --nproc_per_node 2 \
  sparseml.yolov5.train \
  --cfg yolov5x6.yaml \
  --weights "zoo:cv/detection/yolov5-x6/pytorch/ultralytics/coco/pruned75-none" \
  --recipe "zoo:cv/detection/yolov5-x6/pytorch/ultralytics/voc/pruned75_quant-none" \
  --data coco128.yaml \
  --batch-size 2 \
  --gradient-accum-steps 32 \
  --hyp hyps/hyp.VOC.yaml \
  --imgsz 1280 \
  --patience 0 \
  --workers 4