Inference Node — runs on each GPU machine, cuBLAS SGEMM via Panama FFI, pipeline forward pass, gRPC activation transfer