feat: add OrcaRouter (orcarouter/) as a first-class provider

[xilema2]

PR Description

## Summary

This PR adds [OrcaRouter](https://www.orcarouter.ai) — an OpenAI-compatible LLM
meta-router — as a first-class provider in aider, parallel to the existing
OpenRouter integration. Users can run any of OrcaRouter's 150+ upstream models
(OpenAI / Anthropic / Google / DeepSeek / xAI / Qwen / MiniMax / Kimi / Z.ai /
…) with a single API key:

```bash
export ORCAROUTER_API_KEY=<key>
aider --model orcarouter/openai/gpt-4o
aider --model orcarouter/anthropic/claude-opus-4.7
aider --model orcarouter-auto   # alias for the adaptive router

I'm an engineer on the OrcaRouter team.

What's added

• aider/orcarouter.py — OrcaRouterModelManager that fetches the public
  pricing/catalog feed at https://www.orcarouter.ai/api/pricing (no auth) and
  exposes per-model max_input_tokens / max_output_tokens /
  input_cost_per_token / output_cost_per_token in the litellm
  get_model_info shape. Computed from OrcaRouter's published quota formula
  (model_ratio * 2 USD/1M for input, × completion_ratio for output).
  24h on-disk cache at ~/.aider/caches/orcarouter_models.json, mirroring the
  OpenRouter manager.
• aider/models.py —
  • ModelInfoManager delegates to OrcaRouterModelManager for
    orcarouter/-prefixed models when litellm has no built-in info.
  • Model.send_completion auto-routes orcarouter/<vendor>/<model> to
    https://api.orcarouter.ai/v1 via litellm's OpenAI-compatible client,
    using ORCAROUTER_API_KEY from the environment. HTTP-Referer /
    X-Title attribution headers are injected so OrcaRouter's console can
    credit traffic to aider (users can override via
    extra_params.extra_headers).
  • ORCAROUTER_API_KEY added to fast_validate_environment keymap so
    aider's startup env check recognises it.
  • orcarouter-auto alias added to MODEL_ALIASES for OrcaRouter's
    adaptive router.
• aider/website/docs/llms/orcarouter.md — new docs page (parallel to
  openrouter.md) covering setup, the adaptive orcarouter/auto router,
  the tool-calling caveat, and extra_body provider-routing config.
• tests/basic/test_orcarouter.py — unit tests mirroring the OpenRouter
  test pattern: pricing formula on a flat-array payload, pricing formula on
  the wrapped {"data": [...]} payload, empty-result handling, and the
  ModelInfoManager delegation.

Design notes

• No litellm upstream changes needed. The orcarouter/ prefix is rewritten
  to openai/<rest> for the underlying litellm call, with api_base and
  api_key injected, so litellm's existing OpenAI-compatible client handles
  the request. The user-facing model name keeps the orcarouter/ prefix so
  aider's per-provider hooks (alias resolution, metadata lookup) work
  uniformly.
• Closely follows the OpenRouter integration shape so it's easy to review
  side-by-side and maintains the same caching / fallback behaviour.

How verified

• Unit tests added (pytest tests/basic/test_orcarouter.py).
• Manual smoke test against https://www.orcarouter.ai/api/pricing confirms
  the pricing formula matches published per-model rates (e.g.
  openai/gpt-4o: model_ratio=1.25, completion_ratio=4 →
  $2.50 / $10.00 per 1M tokens).
• Linted with the project's pre-commit config (Black / isort).

Happy to iterate on naming, doc tone, or scope. Thanks for maintaining aider!

[CLAassistant]

All committers have signed the CLA.

[xilema2]

Hi @paul-gauthier, gentle ping on this PR. As a first-time contributor, the required CI workflows are waiting for maintainer approval to run (4 workflows pending). Could you kindly approve them when you have a moment? Happy to address any feedback. Thanks!