vLLM issue #53328: prepare_nvfp4_moe_layer_for_marlin warns 'Your GPU does not have native support for FP4' when the checkpoint is merely weight-only

Demonstrated on Jetson AGX Thor (SM110) with torch.ops._C.cutlass_scaled_mm_supports_fp4(110) returning True. An identical unconditional warning also exists in compressed_tensors_moe_w4a4_mxfp4.py.