Semantic Ensemble Loss and Latent Refinement for High-Fidelity Neural Image Compression