Guarantees for Nonlinear Representation Learning: Non-identical Covariates, Dependent Data, Fewer Samples