feat(teleop): multi-episode recording without restarting the simulator

[2047767028-lang]

Motivation

autoteleop.sh currently supports exactly one episode per launch:
pressing y/N kills every teleop process (simulator included) and the
script exits. Collecting N episodes means restarting all four terminals
N times, paying the full Isaac Sim startup cost per episode.

This implements the multi-episode half of #181.

How it works

A new /sim/stop_episode topic carries an end-of-episode signal:

• autoteleop.sh: y/n now end only the current episode
  (keep / mark-discarded), print a completeness check of the produced
  files (.mcap, metadata.yaml, recording_info.json), and keep
  waiting — the next record-button press starts the next episode.
  A new q key stops everything (the previous y/N behaviour).
• api_core.py (_on_recording): on the stop signal, the current ros2
  bag is finalized gracefully via SIGINT (so metadata.yaml is
  written), recording_info.json is written if the episode ended
  before frame 100, and the one-shot wait_recording latch is re-armed
  for the next record-button press.
• server_node.py / ros_nodes.py / teleop.py: subscribe to the
  stop signal and reset the cached recording state on both the
  simulator and teleop sides, so the next episode always requires a
  fresh record-button press and never auto-starts from stale signals.

Testing status

Being transparent here: the logic was developed and validated
end-to-end on the pre-3.2.0 layout (a v3.0-era container) — we used it
to collect multi-episode teleop datasets in one sitting, which later
trained successfully. This PR ports it 1:1 to the current package
layout (insertion points verified against current main; bash -n and
py_compile pass), but our local environment cannot run the 3.2.0
stack, so it has NOT been run end-to-end on 3.2.0 itself. Happy to
adjust anything that doesn't match the new runtime.

🤖 Generated with Claude Code