feat(rl): PPO training pipeline with smoke-train verification

[thanhndv212]

M2b — RL Training Pipeline (COMPLETES M2)

New: training/ package

File	Purpose
train_ppo.py	SB3 PPO entry point (CLI, SubprocVecEnv, VecNormalize)
eval_policy.py	Deterministic evaluation → pos_rmse, success_rate, crash_rate, mean_reward
configs/ppo_hover.yaml	YAML hyperparameters

New examples

• examples/07_rl_hover.py — end-to-end: train PPO 50k steps, eval, print RMSE
• examples/08_rl_vs_pid.py — head-to-head comparison vs cascaded PID on circular trajectory

pyproject.toml — optional RL extras

[project.optional-dependencies]
rl  = ["torch>=2.2", "stable-baselines3>=2.3", "gymnasium>=0.29", "tensorboard>=2.15", "pyyaml>=6.0"]
rl-dev = ["drones-sim[rl]", "wandb", "optuna>=3.5", "moviepy"]

Smoke-train verified ✅

• 50,000 timesteps in 25 s at ~2,000 fps (CPU, 2 envs)
• Model checkpoint saved to training/checkpoints/final.zip
• python -m training.eval_policy --path training/checkpoints/final.zip loads + evaluates
• Pipeline: train → save → load → eval — verified end-to-end

Tests — 2 new in tests/test_rl_training.py

• Smoke-train 1000 steps + save/reload/predict
• Self-contained eval metrics test (tiny model → metrics dict)
• All guarded by @pytest.mark.skipif(not _HAS_SB3)

.gitignore

Added training/checkpoints/ and tb/ (TensorBoard logs)

============================== 92 passed ==============================

---

M1 + M2 — COMPLETE 🎉

6 PRs, 35+ files, 2,000+ lines of code, 92 tests passing:

1. ✅ Quaternion dynamics plant (no gimbal lock)
2. ✅ CI workflow (ruff + pytest, Python 3.10–3.12)
3. ✅ Disturbance framework (wind, ground effect, payload drop, motor failure)
4. ✅ Telemetry logging (CSV + JSON Lines)
5. ✅ RL Gymnasium env (modular action/obs/reward/task)
6. ✅ RL training pipeline (PPO smoke-trained, verified)