Finite-Time Last-Iterate Convergence for Multi-Agent Learning in Games