feat: Reproduce VisDrone + SKU-110K baselines for Issue #49

[dcdc4747]

Summary

Reproduce YOLO-Master-v0.1-N and YOLO-Master-EsMoE-N baselines on VisDrone and SKU-110K datasets as required by #49.

Hardware: RTX 4080 Laptop GPU (12GB), Windows 11, PyTorch 2.5.1+cu124

Results

Dataset	Model	mAP50	mAP50-95	box_loss	cls_loss	moe_loss
VisDrone	YOLO-Master-v0.1-N	0.3044	0.1712	1.388	0.936	1.000
VisDrone	YOLO-Master-EsMoE-N + dense eval	0.3098	0.1744	1.397	0.952	1.000
SKU-110K	YOLO-Master-v0.1-N	0.8827	0.5415	1.301	0.553	1.000
SKU-110K	YOLO-Master-EsMoE-N + dense eval	0.8835	0.5442	1.305	0.550	1.000

⚠️ EsMoE-N requires --dense-eval-for-esmoe — without it VisDrone mAP50 collapses from 0.31 to <0.10.

W&B

https://wandb.ai/acjojo2-communication-university-of-china-nanjing/yolo_master_issue49

Training Commands

# VisDrone
python scripts/issue49/yolo_master_issue_49.py --dataset VisDrone --model YOLO-Master-v0.1-N --batch 12 --epochs 100 --workers 2
python scripts/issue49/yolo_master_issue_49.py --dataset VisDrone --model YOLO-Master-EsMoE-N --batch 12 --epochs 100 --workers 2 --dense-eval-for-esmoe

# SKU-110K
python scripts/issue49/yolo_master_issue_49.py --dataset SKU-110K --model YOLO-Master-v0.1-N --batch 13 --epochs 100 --workers 2
python scripts/issue49/yolo_master_issue_49.py --dataset SKU-110K --model YOLO-Master-EsMoE-N --batch 13 --epochs 100 --workers 2 --dense-eval-for-esmoe

Changes

- Fix METRIC_COLUMNS in issue49 script (mixture_aux_loss instead of moe_loss)
- Add reproduction pipeline script run_all_issue49.sh
- Update README with reproduction results and hardware notes
- Add summary CSVs for all 4 experiments

Closes #49