This PR makes sure timers are finalized before clocks. Since a timer uses a clock, the clock must be valid longer than the timer.

I suspect this will fix the windows test failures in ros2/rcl#286. That PR adds a time jump callback to the clock in rcl_timer_init(), and removes it in rcl_timer_fini(). If the clock is finalized before the timer then the attempt to remove the jump callback results in trying to access an rcl_clock_t after it has been freed.

TimerBase already has a shared pointer to Clock for this purpose, but this is insufficient. TimerBase::get_timer_handle() returns an std::shared_ptr<rcl_timer_t> with a deleter that calls rcl_timer_fini(). This handle can live longer than TimerBase in an executor's memory strategy. Specifically AllocatorMemoryStrategy keeps a list of std::shared_ptr<rcl_timer_t>. If the NodeClock and TimerBase are destroyed before an Executor then the clock is finalized before the timer handle which may result in a crash.

CI

• Linux
• Linux-aarch64
• macOS
• Windows

blocks ros2/rcl#286