feat: add LFM2-VL model support

[reacher-z]

Summary

• add a Transformers adapter for the LiquidAI LFM2-VL model family
• register the 450M, 1.6B, and 3B checkpoints in supported_VLM
• preserve interleaved image/text inputs and use the official multimodal apply_chat_template path
• support CUDA, MPS, CPU, and an optional device_map
• add focused regression tests for message construction, MM-IFEval ordering, generation slicing, and model placement

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

Implementation follows the official model card inference API: https://huggingface.co/LiquidAI/LFM2-VL-450M

Validation

• python3 -m pytest -q tests/test_lfm2_vl.py — 3 passed
• python3 -m py_compile scieval/config.py scieval/vlm/__init__.py scieval/vlm/liquid.py tests/test_lfm2_vl.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, and the public LiquidAI/LFM2-VL-450M weights: loaded in 60.98s and generated in 4.85s; for assets/radar.png and the prompt “What kind of chart is shown? Answer briefly.” the adapter returned Polar.

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 the tests and real-model smoke inference before submission.