On Nonasymptotic Confidence Intervals for Treatment Effects in Randomized Experiments