Quantile universal threshold: model selection at the detection edge for high-dimensional linear regression