feat: add OrcaRouter as a named VLMEvalKit API provider

[nissrin2020ali-ux]

Motivation

VLMEvalKit evaluates large vision-language models against a large registry of named API providers — Together AI, SiliconFlow, MiniMax, and others each have a first-class wrapper in vlmeval/api/ plus preset entries in vlmeval/config.py. When benchmarking with an external model gateway, users currently have no way to pick OrcaRouter without treating it as an anonymous custom --base-url. This PR adds OrcaRouter as a first-class provider in the same style as the existing Together AI integration, so Cosmos Reasoner benchmarks can run against OrcaRouter's model namespace directly.

OrcaRouter is an OpenAI-compatible AI gateway built for both models and agents. Like OpenRouter, it exposes a provider/model namespace across many models — but it also combines adaptive routing, automatic failover, zero-markup inference, observability, guardrails, and agent-tool governance behind the same endpoint. Adding orcarouter as a first-class provider means this project's users can use that stack directly, without treating OrcaRouter as an anonymous custom base URL.

It also runs gateway-level, zero-trust security for AI agents on the same endpoint — screening every prompt/response and governing every tool call on a default-deny basis, with no application code changes.

Changes

• vlmeval/api/orcarouter.py — new OrcaRouterAPI wrapper, modeled directly on vlmeval/api/together.py: OpenAI-compatible chat completions with image_url (base64 or URL) support, ORCAROUTER_API_KEY env key resolution, retries, and a configurable api_base defaulting to https://api.orcarouter.ai/v1/chat/completions.
• vlmeval/api/__init__.py — import and export OrcaRouterAPI.
• vlmeval/config.py — four preset api_models entries: OrcaRouter_Gemini-3.1-Flash-Lite, OrcaRouter_Gemini-3.5-Flash, OrcaRouter_Claude-Sonnet-5, and OrcaRouter_DeepSeek-V4-Flash-Vision.
• docs/en/Quickstart.md — add ORCAROUTER_API_KEY to the documented .env API-key list.

Verification

• python -m py_compile and flake8 --max-line-length=99 pass on the new files (config.py is pre-commit-excluded).
• L3 live round-trips through the new OrcaRouterAPI class against the live https://api.orcarouter.ai/v1 endpoint succeeded for all four preset models (google/gemini-3.1-flash-lite, google/gemini-3.5-flash, anthropic/claude-sonnet-5, deepseek/deepseek-v4-flash-vision-exp), each returning ret_code=0 with a valid completion.

Discord: discord.gg/YEubt8enRA · X: https://x.com/OrcaRouter

I'm an engineer on the OrcaRouter team.