[Model] Add VisionPsy-Nano-460M / -Flash support

[cdpark92]

What

Adds support for VisionPsy-Nano-460M and VisionPsy-Nano-460M-Flash
(qvac/VisionPsy-Nano-460M,
qvac/VisionPsy-Nano-460M-Flash),
two ~460M-parameter vision-language models built for on-device use
(SigLIP2-base-512 vision tower + SmolLM2-360M decoder, pixel-shuffle
projector, dynamic 512px tiling). The Flash variant keeps the native
resolution with a min-side clamp instead of resizing to the max side, trading
a little accuracy for much lower time-to-first-token.

Changes

All additive — no existing file or adapter is modified.

vlmeval/vlm/visionpsy.py (new) — self-contained adapter with two
backends:

1. transformers (default) — the checkpoints are Hub-packaged and bundle
   their own processor and modeling code, so they load with
   trust_remote_code=True and need no extra install (INSTALL_REQ = False).
2. in-process vLLM (--use-vllm) — serves the same checkpoint through the
   VisionPsy vLLM plugin, which registers the out-of-tree architecture. The
   engine is built before any tokenizer/torch work: HF tokenizers and torch
   spawn worker threads, and vLLM forks its engine-core subprocess, so the
   reverse order deadlocks.

The base/Flash preprocessing variants are resolved from the checkpoint's own
config, so one class serves both models.

vlmeval/vlm/__init__.py (+1) export, vlmeval/config.py (+9)
registry entries, vlmeval/inference.py (+1) forwards use_vllm for
VisionPsy model names.

Usage

# transformers (eager) — no extra install
python run.py --data MMStar --model VisionPsy-Nano-460M

# in-process vLLM
pip install "git+https://github.com/tether-ai-research/qvac-visionpsy-nano#subdirectory=vllm-inference"
python run.py --data MMStar --model VisionPsy-Nano-460M-Flash --use-vllm

Verification

• Both models generate correctly on both backends (transformers and
  --use-vllm), greedy decoding, loaded through the registered names.
• pre-commit run passes on the added/changed files.

Notes

• The models are fine-tuned from
  nanoVLM-460M-8k (MIT), but
  ship as standalone Hub packages with their own processor, so they get their
  own adapter rather than reusing nanovlm.py.
• Loading from a local checkpoint folder requires transformers>=5.14.1
  (fix: include transitive relative imports when loading from local directory huggingface/transformers#46022 fixed transitive relative imports for local
  trust_remote_code loads); the registered Hub ids work on older versions
  as well.
• MM-IFEval evaluation additionally needs the rule-based-verifier import fix
  proposed in [Fix] MM-IFEval: import rule-based verify functions into module globals #1601.

[cdpark92]

Please refer this blog post for more information about this model. Thanks!
https://huggingface.co/blog/qvac/visionpsy