feat(llm): add Atlas Cloud provider preset

[binyangzhu000-sudo]

Summary

• add an atlascloud OpenAI-compatible provider preset at https://api.atlascloud.ai/v1
• register Atlas Cloud in the interactive init flow with isolated ATLASCLOUD_API_KEY configuration
• provide maintained DeepSeek V4 Pro/Flash defaults, a commented config example, and focused provider tests

Validation

• .venv/bin/python -m pytest tests/test_atlascloud_provider.py tests/test_minimax_provider.py tests/test_rc_cli.py -q (51 passed, 3 skipped)
• .venv/bin/python -m pytest tests/ -q (2869 passed, 56 skipped; one existing AsyncMock warning in tests/test_servers.py)
• .venv/bin/python -m compileall -q researchclaw tests/test_atlascloud_provider.py
• verified the configured models and OpenAI chat-completions capability against the public Atlas Cloud model catalog

Build note

pip wheel . --no-deps reaches the existing package-data step, then fails because researchclaw/templates/styles/iclr_2025/iclr2025_conference.bst is added to the wheel twice. This change does not touch packaging or template files.

[Jiaaqiliu]

Merged. One adjustment during integration, since #303 landed just before this and also claimed a wizard slot.

#303 added three MiniMax regional presets at choices 7/8/9, which collided with atlascloud at 7. Atlas Cloud is now choice 10.

Rather than re-pin the number, I made the test look it up:

choice = next(
    key for key, value in _PROVIDER_CHOICES.items() if value[0] == "atlascloud"
)
monkeypatch.setattr("builtins.input", lambda _prompt: choice)

test_init_wizard_writes_atlascloud_defaults hardcoded "7", so it would have kept passing while driving the wizard into MiniMax — asserting on the wrong provider's output. The next preset PR would have hit the same trap. Everything else merged unchanged.

Verified post-merge: tests/test_atlascloud_provider.py + test_minimax_provider.py + test_rc_cli.py → 64 passed, 3 skipped. Full suite: 2956 passed.