Learning a Class of Mixed Linear Regressions: Global Convergence under General Data Conditions