Convergence of the Backward Deep BSDE Method with Applications to Optimal Stopping Problems