Run independent affine M=1 matrix-vector products in one native batched call.
Weight/scales/bias views share their original buffers. MLX 0.32.2's
flattened M>=2 path uses different arithmetic, so appending known tokens
must explicitly select this layout when preserving decode state.
Run independent affine M=1 matrix-vector products in one native batched call. Weight/scales/bias views share their original buffers. MLX 0.32.2's flattened M>=2 path uses different arithmetic, so appending known tokens must explicitly select this layout when preserving decode state.