Near Optimal Heteroscedastic Regression with Symbiotic Learning