Pushing the Limits of BFP on Narrow Precision LLM Inference