feat(cosim): optional seconds_stopped in set_agent_states

[01Dami23]

What:

• Add an optional seconds_stopped array to c_set_agent_states (+ its binding and the Drive.set_agent_states wrapper) that injects each agent's stopped-time as external state. NULL / None keeps c_step's own per-tick accumulation.
• Expose AGENT_STOPPED_SPEED_THRESHOLD to Python (mirrors MAX_SPEED / ACCEL_LONG_NORM) so co-sim callers read the threshold instead of hardcoding it when computing seconds_stopped.
• Unit test covering both branches (NULL preserves, array overwrites).

Why:

• When a co-sim overwrites an agent's kinematics every tick instead of letting PufferDrive simulate it, c_step derives seconds_stopped from the dummy/injected rollout speed at env.dt, not the true external speed at the real tick rate. For CARLA/nuplan the ego is PufferDrive-simulated (only background is injected), so their ego is unaffected; a caller that injects the ego itself needs to own the field. Injecting it also removes the partner/parked-slot staleness (the existing TODO(hack)).

Notes:

• Backward compatible: every current caller passes no seconds_stopped (Python default None -> C NULL), so behavior is unchanged.