feat: llamafile_sgemm bias support

[chenghuaWang]

Tested on X86 and Arm device using Qwenv1.5 model.

For further optimisation or neatness of gemm code: we can concat ones at the last col of $$X$$ and concat $$\text{Bias}$$ at the last row of $$W$$ when allocating Tensor if they have same precision.