The "good" Boussinesq equation: a Riemann-Hilbert approach