Block Coordinate Descent for Deep Learning: Unified Convergence Guarantees