Beyond Filtering: Adaptive Image-Text Quality Enhancement for MLLM Pretraining