Gravar-mail: Evaluating the Impact of Large Language Model-Generated Synthetic Data on Recommender Systems Performance