On assumption-free tests and confidence intervals for causal effects estimated by machine learning