Semantics Alignment via Split Learning for Resilient Multi-User Semantic Communication