Optimize the concat and split cuda implementation for cases when the number of inputs/outputs is less than 5. (#17979)

test=develop