Exact minimax risk for linear least squares, and the lower tail of sample covariance matrices