serve: GPU-batched, CUDA and APR/SafeTensors GPU routers are built without AuthGate (APR_API_KEY_HASH ignored) #5478

Found by the #5451 quorum. The CPU GGUF router is gated now; these still build create_router_with_config with no auth layer: crates/apr-cli/src/commands/serve/server.rs (start_gguf_server_gpu_batched), handler_gpu_completion.rs (start_gguf_server_cuda, past its early serve_router fallback), handlers_include_01.rs (start_apr_server_gpu, start_apr_q4k_server_gpu, start_safetensors_server_gpu). Contract: contracts/apr-serve-api-key-auth-v1.yaml (health/metrics not exempt). Fix: route each through AuthGate::from_env() + auth::layer, with a 401/200 test per router (FALSIFY-AUTH-001).