Add VSI-SUPER benchmark

[akawincent]

Summary

• add VSI-SUPER Recall task configs and scoring utilities
• add VSI-SUPER Count task configs and scoring utilities
• add VSI-SUPER Count Streaming task configs and scoring utilities

Dataset pages:

• https://huggingface.co/datasets/nyu-visionx/VSI-SUPER-Count
• https://huggingface.co/datasets/nyu-visionx/VSI-SUPER-Recall

In scope

• add vsisuper/recall task configs and vsr_utils.py
• add vsisuper/count task configs and vsc_utils.py
• add vsisuper/count_streaming task configs and vsc_streaming_utils.py

Out of scope

• model integrations
• dataset content changes
• changes outside the VSI-SUPER task family

Validation

• pre-commit run --files lmms_eval/tasks/vsisuper/count/vsc_utils.py lmms_eval/tasks/vsisuper/count_streaming/vsc_streaming_utils.py lmms_eval/tasks/vsisuper/recall/vsr_utils.py | result: pass

Risk / Compatibility

• low risk: changes are limited to new task configs and task-local utilities
• no intended behavior changes outside the new VSI-SUPER tasks

Type of Change

• Bug fix (non-breaking change)
• New feature
• New benchmark/task
• New model integration
• Breaking change
• Documentation update
• Refactoring (no functional changes)

[Luodian]

LGTM — clean additive benchmark. Task configs and utilities are well-structured across recall, count, and count_streaming variants.