fix: align timeout bootstrap with final observation contract

[TATP-233]

Fixes #236

Summary

• align synchronous PPO timeout bootstrap with the env final_observation contract instead of reusing V(obs_t)
• keep APPO and off-policy collector actor inputs on the post-reset observation while still storing true terminal observations for learning targets
• move final_observation handling behind a shared transition-bootstrap contract instead of duplicating worker-side patching rules

Impact

• PPO / RSL-RL
• MLX PPO
• APPO collector / learner
• off-policy collector

Validation

• UV_CACHE_DIR=/tmp/uv-cache make check
• UV_CACHE_DIR=/tmp/uv-cache uv run pytest tests/utils/test_final_observation.py tests/algos/test_rsl_rl_ppo.py tests/algos/test_appo_worker.py tests/config/test_locomotion_params.py -q
• UV_CACHE_DIR=/tmp/uv-cache uv run pytest tests/utils/test_final_observation.py tests/algos/test_rsl_rl_ppo.py tests/algos/test_appo_worker.py tests/config/test_locomotion_params.py tests/scripts/test_train_scripts.py -k 'test_play_wrapper_step_exports_timeout_bootstrap_obs or test_play_wrapper_policy_obs_mode_actor or test_play_wrapper_uses_current_reset_contract or test_ppo_hydra_default_algo_log_name' -q

Notes

• tests/scripts/test_train_scripts.py -k test_mlx_time_limit_bootstrap_values_use_final_observation still aborts locally because importing scripts/train_mlx_ppo.py triggers a native mlx extension crash in this machine's environment.