Modernize cloud and Ollama model settings

[dovvnloading]

What changed

• remove hardcoded Ollama runtime defaults and migrate legacy sessions to explicit Auto/Inherit/Explicit assignments
• add a provider-neutral model catalog and deterministic runtime resolver
• discover installed Ollama models asynchronously with health and descriptor metadata
• add searchable model pickers and clearer task routing controls
• make cloud model discovery asynchronous with inline status and isolate saved model profiles by provider
• commit API settings only after provider initialization succeeds
• remove silent Gemini model fallbacks and preserve custom IDs as visible, unverified selections

Why

The previous settings flow treated model IDs as opaque strings, blocked the UI during discovery, and could activate models a user did not have installed. Cloud provider switches could also reuse incompatible saved model IDs. This PR makes readiness and selection state explicit while preserving compatibility adapters for existing runtime callers.

Validation

• python -m compileall -q graphlink_app
• git diff --check
• headless Qt settings construction smoke test
• python -m pytest -q — 414 passed, 1 warning

The internal proposal remains local-only under doc/ and is excluded from this PR.