docs(reference): explicit specs for the grouped-GEMM kernels
The batched grouped kernels (triple/quad/dual_gemm) share their math
with the existing shape-agnostic references, but the mapping was not
nameable. Add explicit reference aliases and rewire the kernel test to
anchor on the reference module instead of inline F.linear math.
Signed-off-by: Guokai Ma <guokai.ma@intel.com>
