On Separation Between Best-Iterate, Random-Iterate, and Last-Iterate Convergence of Learning in Games