E-QRGMM: Efficient Generative Metamodeling for Covariate-Dependent Uncertainty Quantification