[Benchmark] Add support for Ref-L4_test benchmark

[rshube]

Summary

This PR adds support for the Ref-L4_test benchmark in VLMEvalKit.

Changes

• adds a dedicated RefL4Dataset
• registers Ref-L4_test in the dataset registry
• keeps Ref-L4_test separate from RefCOCO
• adds a dedicated Ref-L4 evaluator instead of reusing the plain RefCOCO summary unchanged

Dataset format

Expected TSV columns:

• index
• image
• question
• answer
• bbox_x1
• bbox_y1
• bbox_x2
• bbox_y2
• width
• height
• bbox_area
• bbox_id
• ori_category_id
• image_id
• file_name
• is_rewrite

Evaluation

The evaluator reports Ref-L4-style grounding metrics, including:

• annotation-level accuracy at IoU 0.5
• annotation-level accuracy at IoU 0.75
• annotation-level accuracy at IoU 0.9
• annotation-level mAcc over 0.5:0.95
• size-level metrics
• class-level average metrics

It also supports bbox predictions in both:

• xyxy
• xywh

Notes

• this PR only adds benchmark support code
• hosted DATASET_URL / DATASET_MD5 can be filled once the final TSV hosting location is settled (sending email to maintainers)