Text this: Domain-Incremental Learning Paradigm for scene understanding via Pseudo-Replay Generation