Linear $Q$-Learning Does Not Diverge: Convergence Rates to a Bounded Set