Feature/gemmbf16xfp32

[shaochangxu]

This mr addresses the performance bottleneck of BF16xFP32 matrix multiplication (GEMM) in LLM inference and proposes a high-precision FP32 GEMM operator based on double BF16 residual decomposition. In traditional approaches, naive FP32 GEMM relies on CUDA Cores, achieving high accuracy but low throughput, while reduced-precision schemes such as BF16 or TF32 can leverage Tensor Core acceleration but suffer from severe accuracy loss. The proposed method decomposes the FP32 weight matrix W into a high‑part BF16 tensor W_high and a low‑residual BF16 tensor W_low, sets the scaling factor scale = 1/256, and equivalently transforms the original GEMM into a linear combination of two BF16 GEMMs.