Compute the number of training steps per epoch from dataset size and batch size, for training-loop and sharding checks.
分散式下每卡步數不變,總樣本覆蓋相同。