Towards a Unified Framework for Uncertainty-aware Nonlinear Variable Selection with Theoretical Guarantees