Decouple gateway managers from rclcpp via provider injection for testability and reduced flakiness

[bburda]

Summary

Today every gateway test pulls rclcpp and the ament environment because the manager classes (DataAccessManager, OperationManager, ConfigurationManager, FaultManager facade, LogManager, TriggerManager) take rclcpp::Node * directly and route I/O inline. That coupling makes unit tests heavy (multi-second startup), forces every test fixture through MultiThreadedExecutor lifecycle, and makes test failures timing-sensitive - flaky process spawned but graph not yet seen and subscription destroyed mid-callback patterns appear repeatedly in CI history.

This issue tracks the next pass after #391: each manager keeps its routing and business logic but gains a register_provider(...) injection point, with the ROS-specific I/O extracted into Ros2*Provider classes living in the adapter layer. After the change, manager unit tests compose with mock providers and link only against gateway_core (the neutral library introduced in #391), which removes rclcpp from the unit-test link line and eliminates the executor-lifecycle setup that is the root cause of intermittent failures.

---

Proposed solution (optional)

Concretely:

• Six managers gain provider injection: Data, Operation, Configuration, Fault, Log, Trigger. Each manager moves to core/managers/. The ROS-specific default behaviour is extracted into Ros2*Provider implementations under src/ros2/providers/ (or equivalent path in the adapter layer), registered statically by gateway_node at startup. Plugins continue to register per-entity providers via PluginManager exactly as today.
• RuntimeDiscoveryStrategy becomes Ros2RuntimeIntrospection : IntrospectionProvider. The discovery framework in core/ then routes through the existing IntrospectionProvider chain and stops referencing rclcpp::Node. HybridDiscoveryStrategy collapses into the standard MergePipeline configuration.
• TypeIntrospection moves to ros2_medkit_serialization, where rosidl_typesupport_cpp and rosidl_typesupport_introspection_cpp already live. The gateway then depends on the serializer for type schemas instead of duplicating the rosidl bridge inside the gateway tree.
• Runtime discovery trigger switches from cyclic wall_timer(refresh_interval_ms_) polling to event-driven graph_event->check_and_clear(). The polling timer remains as a low-frequency safety backstop. Reduces idle CPU when the graph is stable and improves new-node detection latency.

Out of scope (intentionally deferred): folding core/ into a separate colcon package - that is a layout change and does not affect the test-link-line goal.

---

Additional context (optional)

Outcome metrics expected after the work:

• Manager unit tests link only gateway_core + GTest (no ament_target_dependencies), proven by extending test_gateway_core_smoke to include a manager headers and have a representative manager test compile against the mock-provider variant.
• Idle gateway process CPU usage drops measurably once the discovery refresh stops firing on a stable graph (sample with top before/after on the demo nodes scenario).
• Existing test suite (~2500 unit, ~3200 integration, ~2600 clang-tidy) stays green without timing-sensitive sleeps in tests we touch.
• New-node detection latency (time from ros2 run to gateway /apps listing the new entity) drops below 500 ms in steady-state operation.