Optimal Low-Rank Matrix Completion: Semidefinite Relaxations and Eigenvector Disjunctions