feat(models): add NanoVLM chat model with async multi-GPU eval

[Jinghao-Guo]

Summary

• Add NanoVLM evaluation model (lmms_eval/models/chat/nanovlm.py) as a chat-style model for lmms-eval
• NanoVLM is a lightweight VLM architecture (SigLIP2 + MLP projector + Qwen3-0.6B) trained with lmms-engine
• Supports async multi-GPU inference following the async_hf_model pattern: loads model replicas on N GPUs, dispatches work via job queue with independent worker threads
• Handles NanoVLM-specific image token expansion (<|image_pad|> -> 256 tokens)

Usage

# Single GPU
python -m lmms_eval --model nanovlm --model_args pretrained=LMMs-Lab-Speedrun/NanoVLM_Init --tasks mme --batch_size 1

# Multi-GPU async (auto-detects all visible GPUs)
CUDA_VISIBLE_DEVICES=0,1,2,3 python -m lmms_eval --model nanovlm --model_args pretrained=LMMs-Lab-Speedrun/NanoVLM_Init --tasks mme --batch_size 1

# Explicit GPU selection
python -m lmms_eval --model nanovlm --model_args "pretrained=LMMs-Lab-Speedrun/NanoVLM_Init,worker_gpus=0,1,2,3" --tasks mme --batch_size 1

Test plan

• Verified single-GPU eval on MME benchmark (2374 samples)
• Verified async 4-GPU eval on full SpeedRun benchmark suite (11455 samples across 6 benchmarks)
• Pre-commit hooks (black, isort) pass