Polylogarithmic width suffices for gradient descent to achieve arbitrarily small test error with shallow ReLU networks