On the convergence properties of a $K$-step averaging stochastic gradient descent algorithm for nonconvex optimization