Add a sub-45 BC student controller

[teerthsharma]

What changed

This PR adds bc_student, a compact behavioral-cloning controller backed by a Torch MLP and a small saved checkpoint.

The short version: the earlier PID/topology experiments got close enough to be educational, but not close enough for the leaderboard cutoff. I switched to the idea that has clearly been winning this challenge: optimize good actions offline, then distill that behavior into a fast student policy. The resulting controller keeps the runtime simple: build the same state/future/past-action feature vector every step, run one CPU Torch forward pass, and clip the steering command into the simulator range.

Score

Full 5,000 segment local evaluation:

controller   lataccel_cost   jerk_cost   total_cost
bc_student   0.525043        17.537342   43.789511

That is below the strict <45 target and below the <50 range requested for leaderboard consideration.

Why this feels worth submitting

I did not want to send another clever-looking controller that only wins a small smoke test. This one got the full 5k pass before opening the PR. It is more direct, less hand-tuned, and much closer to the current leaderboard style: let offline trajectory optimization teach a policy, then submit the policy as the controller.

Validation

python -m py_compile controllers/bc_student.py
pytest -q test_bc_student.py
python -u eval_segment_action_replay.py --model_path ./models/tinyphysics.onnx --data_path ./data --num_segs 5000 --controller bc_student --workers 4 --chunksize 10 --output ./runs/bc_student_eval_5000.csv

The committed test covers model loading, finite/clipped actions, short future-plan padding, and model-cache reuse.

Notes

The controller depends on torch>=2.0 for inference. The checkpoint is intentionally included so the controller is self-contained for evaluation.

[teerthsharma]

Superseded by #41 from the renamed Banach branch.