Correcting Mean Bias in Text Embeddings: A Refined Renormalization with Training-Free Improvements on MMTEB