Tools: ros2: tear down rclpy cleanly in the AP_DDS tests

[peterbarker]

Summary

Fixes segfault recently starting in colcon tests

Classification & Testing (check all that apply and add your own)

• Checked by a human programmer
• Non-functional change
• No-binary change
• Infrastructure change (e.g. unit tests, helper scripts)
• Automated test(s) verify changes (e.g. unit test, autotest)
• Tested manually, description below (e.g. SITL)
• Tested on hardware
• Logs attached
• Logs available on request

Description

Every test in ardupilot_dds_tests tore down with nothing but rclpy.shutdown(), which shuts the context down while that context's node, subscriptions and DDS participant are all still alive, and left its rclpy.spin() thread unjoined. colcon runs the whole package in one interpreter, so those zombie endpoints accumulated from test to test.

They race with Fast-DDS intra-process delivery, and the process intermittently died with SIGSEGV on an internal DDS thread:

History::get_change_nts
StatefulReader::NotifyChanges
StatefulReader::change_received
StatefulReader::processDataMsg
StatefulWriter::deliver_sample_to_intraprocesses
FlowControllerImpl<...>::run

Both CI occurrences were in the 22nd of the 23 tests, i.e. at maximum accumulation, and killed the whole pytest run with exit code -11.

rclpy.spin() cannot be torn down cleanly: it uses a process-wide global executor bound to whichever context was current when the first test span, and the only way to make it return is to shut the context down. So spin on a private SingleThreadedExecutor via a ros_node() context manager, which shuts the executor down, joins the spin thread and destroys the node while the context is still valid.

Reproduced by repeating the time tests in one pytest process (pytest --count=40). Over 6 shards x 4 loops, with identical harness and only this change differing: 4 of 24 loops segfaulted before, 0 of 24 after, with all 80 tests passing in every loop.