[Pranav-Singh-Devloper]

Fixes #306

The user process was started with stderr=sys.stdout, which is RAM's own stdout (the container log). In Python exercises stderr only reaches the console after import WebGUI, because start_console() replaces sys.stderr there, so anything raised before that import was lost. With stderr=subprocess.STDOUT the process's stderr goes to the same console as its stdout. The C++ templates already reopen stderr on the console, so Python now behaves the same way.

Testing (Image Classification):

• Wrong model file name before import WebGUI: the NoSuchFile traceback is now shown in the console (before: empty console).
• Working code runs normally and the GUI works.
• An exception raised after import WebGUI is shown once, not twice.
• import onnxruntime, torch, cv2 and rclpy print nothing to stderr in the RADI, so this adds no noise.
• black --check passes.

Please merge together with JdeRobot/RoboticsAcademy#4101. Without it, every run shows two Error sending message: socket is already closed. lines from gui_interfaces in the console.

Before

ram-stderr-before.mp4

After

ram-stderr-after.mp4

[Pranav-Singh-Devloper]

The failing test check isn't caused by this change. All 21 failures are KeyError: 'scene' at manager.py:325 (cfg_dict["scene"]), because the test configs don't include a scene key. The same tests fail the same way on #305 (run). I can open a separate issue for the outdated tests if that's useful.