vLLM merges fix for CohereASR FP16 attention bias dtype

The vLLM project has integrated a fix for attention bias data type issues when running CohereASR in FP16 precision. This ensures improved numerical stability and accuracy for Cohere's automatic speech recognition models in the inference runtime.