On Bellman equations for continuous-time policy evaluation I: discretization and approximation