Implement Autoregressive Multi-Track Finder

[dhshin04]

Summary

• Adds models/AutoregressiveTrackFinder.py: inference-only iterative loop that reuses the existing single-track TrackFinder — predict one dimuon pair, remove its hits from the matrix, repeat until confidence drops below threshold or max_iters is reached; only runs inference on still-active events each iteration
• Adds eval_autoregressive.py: evaluation using Hungarian matching between predicted and GT pairs (in confidence order, not canonical order), with per-detector residuals, chi-squared, and TP/FN/FP per iteration
• Adds scripts/eval_autoregressive.slurm: SLURM job for baseline inference + evaluation on the HPC cluster
• Adds data/multi_track/gen_residual_training.py (optional): generates single-track format training data from multi-track events with progressive hit removal, for fine-tuning
• Adds scripts/train_autoregressive.slurm (optional): SLURM job to generate residual data, fine-tune, and evaluate

No new model architecture — the baseline reuses track_finder_64.keras as-is.

Test plan

• Smoke test: compute_confidence and remove_predicted_hits unit tests pass (uniform confidence ≈ 1/201, peaked ≈ 1.0, hit removal correct)
• Ruff linting passes on all new Python files
• Run eval_autoregressive.slurm on HPC with multi-track validation data; verify .npz output and evaluation metrics print correctly
• (Optional) Run train_autoregressive.slurm to validate the fine-tuning pipeline end-to-end