feat: add nanoVLM model support

[reacher-z]

Summary

• add a SciEvalKit adapter for the official pure-PyTorch nanoVLM implementation
• register the public 230M-8k and 460M-8k checkpoints
• support multiple images, current image-splitting checkpoints, text-only inputs, and automatic CUDA/MPS/CPU selection
• provide an actionable installation error directing users to set NANOVLM_PATH
• add focused regression tests for model setup, image preprocessing, prompt construction, generation, decoding, and text-only input

This is an independent new foundation-model integration, one of the major-contribution categories described in the SciEvalKit README.

Sources:

• official implementation: https://github.com/huggingface/nanoVLM
• public checkpoint: https://huggingface.co/lusxvr/nanoVLM-230M-8k

Validation

• python3 -m pytest -q tests/test_nanovlm.py — 2 passed
• python3 -m py_compile scieval/config.py scieval/vlm/__init__.py scieval/vlm/nanovlm.py tests/test_nanovlm.py
• pre-commit hooks pass for the new adapter, import, and tests
• real MPS smoke test against the current official nanoVLM checkout and public lusxvr/nanoVLM-230M-8k weights: the adapter loaded a cached checkpoint in 8.89s, generated in 3.67s, and returned Cat. for the official cat image and prompt “What animal is shown? Answer briefly.”

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.

Setup

git clone https://github.com/huggingface/nanoVLM.git
export NANOVLM_PATH=/path/to/nanoVLM

Install the dependencies listed by nanoVLM, then select nanoVLM-230M-8k or nanoVLM-460M-8k through SciEvalKit.

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.