Robust bilinear factor analysis based on the matrix-variate $t$ distribution