Amortized Inference of Multi-Modal Posteriors using Likelihood-Weighted Normalizing Flows