Sharp high-probability sample complexities for policy evaluation with linear function approximation