feat(rl-env): populate observation with task instruction

[abrichr]

WAALiveAdapter._get_observation() does not populate raw_observation with the task instruction. RLEnvironment stores _current_task.instruction but never exposes it on observations.

This means RL agents using observations cannot see what task they should perform.

Fix: populate raw_observation["instruction"] in observe(), or add a public current_instruction property to RLEnvironment.