Self-Normalized Inference for Constant-Stepsize Temporal-Difference Learning under Markovian Sampling