What is the transaction ID wraparound problem in PostgreSQL, why is it critical for large databases, and what role does the VACUUM process play in preventing it?