fix: assert float8_e4m3fn dtype in _prepack_weight_sf

Non-prepacked path handles uint32 unpacking, but prepack sends
weight_sf directly to _C.prepack_sfb(). If scales aren't float8,
this would silently produce wrong layout/values.

src/nvfp4_megamoe_kernel/nvfp4_mega_moe.py

@@ -109,6 +109,8 @@ def _prepack_weight_sf(weight_sf, N, K, tag):
     if cache_key in _prepack_weight_sf._cache:
         return _prepack_weight_sf._cache[cache_key]
 
+    assert weight_sf.dtype == torch.float8_e4m3fn, weight_sf.dtype
+
     from nvfp4_megamoe_kernel.cutlass_nvfp4_gemm.kernel import prepack_sfb
 
     E = weight_sf.shape[0]