Expand description
MarlinExpertStack<CpuBackend> impl on top of CPU’s dequant-on-load
GptqStore. Facade — delegates to the existing
BackendQuantMarlin::moe_gemm_phase_* (default trait impl that loops
calling gemm_gptq_with_offset_strided on CPU) and
make_stacked_expert_linear methods.
CPU has no real Marlin tiles; the impl exists so the bucketed MoE
path’s parity test (tests/moe_bucketed_parity_test.rs) still
compiles after the Phase C trait-object migration. Phase C step 4
inlines the kernel calls here and deletes the trait methods.