Extreme Compression of Per-Layer Embeddings via Cross-Layer Latent Vector Quantization