Honor catalog-declared model setting choices

[breedx]

Summary

• prefer a model catalog's setting_choices for choice-based settings
• retain the existing reasoning capability flags as the fallback
• ignore unknown advertised values by intersecting with client-supported choices

This lets remotely discovered models expose valid levels such as xhigh without hardcoding model IDs.

Verification

• pytest -q --no-cov tests/command_line/test_model_settings_menu_coverage.py — 93 passed
• Ruff check and format clean for the two changed files