Text this: Versioned Late Materialization for Ultra-Long Sequence Training in Recommendation Systems at Scale