feat: Add Spatial-DISE benchmark task

[shinmohuang]

Summary

• Add Spatial-DISE (ICLR2026) as a new lmms-eval task for 2D/3D visual-spatial reasoning.
• Load the official 559-example benchmark split from TACPS-liv/Spatial-DISE.
• Evaluate A/B/C/D multiple-choice predictions with exact answer-letter extraction.

In scope

• Adds lmms_eval/tasks/spatial_dise/spatial_dise.yaml.
• Adds lmms_eval/tasks/spatial_dise/utils.py.
• Reads Spatial-DISE images from Hugging Face tar shards and maps CSV images/... paths to tar member paths.

Out of scope

• Does not modify existing lmms-eval task behavior.
• Does not add or change model integrations.
• Does not change Spatial-DISE data, labels, splits, or evaluation semantics.

Validation

• python3 -m py_compile lmms_eval/tasks/spatial_dise/utils.py && git diff --check && black --check lmms_eval/tasks/spatial_dise/utils.py | sample size: N/A | key metrics: syntax/style/whitespace checks | result: pass
• TaskManager(verbosity='ERROR') discovery check | sample size: N/A | key metrics: spatial_dise registered=True | result: pass
• python -m lmms_eval eval --model dummy --tasks spatial_dise --limit 1 --batch_size 1 | sample size: N=1/559 | key metrics: spatial_dise_acc=0.0 with dummy model | result: pass

Risk / Compatibility

• Low risk: this PR only adds a new task under lmms_eval/tasks/spatial_dise.
• Requires access to the Spatial-DISE Hugging Face dataset tar shards; users may set SPATIAL_DISE_ROOT to reuse a local dataset checkout.

Type of Change

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