feat: add Granite Docling model support

[reacher-z]

Summary

• add a Transformers adapter for IBM Granite Docling 258M document conversion
• register granite-docling-258M in supported_VLM
• preserve image/text ordering and return DocTags without stripping structural special tokens
• support CUDA, MPS, CPU, custom model-loading kwargs, and optional device_map
• add focused regression tests for multimodal message construction, model loading, device placement, generation slicing, and DocTags decoding

This is a third independent new foundation-model integration, one of the major-contribution categories described in the SciEvalKit README. It is separate from the LFM2-VL and nanoVLM integrations in #11 and #12.

Official model card and inference API: https://huggingface.co/ibm-granite/granite-docling-258M

Validation

• python3 -m pytest -q tests/test_granite_docling.py — 2 passed
• python3 -m py_compile scieval/config.py scieval/vlm/__init__.py scieval/vlm/granite_docling.py tests/test_granite_docling.py
• pre-commit hooks pass for the new adapter, import, and tests
• real CPU smoke test with transformers==4.57.1, torch==2.8.0, the public 258M weights, and the official sample arXiv page: the adapter loaded a cached checkpoint in 10.69s, generated 8 tokens in 40.53s, and returned the expected DocTags prefix <doctag><page_header><loc_100><loc
• a longer uncached official-API run generated <doctag><page_header>...Energy Budget of WASP-121 b..., confirming page-content extraction

The full config file has pre-existing lint violations outside this diff, so validation also included git diff --check and compilation of the modified config.

AI assistance disclosure

OpenAI Codex assisted with implementation, tests, and PR drafting. I ran and inspected both the focused tests and real-model smoke inference before submission.