Allow custom Auto model-routing preferences

## Feature request

Allow users to configure task-aware preferences for **Auto** model selection in the GitHub Copilot app.

For example:

- Prefer GPT-5.6 Terra for straightforward coding tasks, focused edits, simple questions, and routine explanations.
- Prefer GPT-5.6 Sol for architecture/design work, complex debugging, security review, multi-file refactors, and other involved changes.

These preferences should be advisory rather than absolute: the router may still choose another enabled model when availability, reliability, cost, or enterprise policy requires it.

Custom instructions currently affect response behavior but do not influence the model selected by Auto. A dedicated setting would make Auto routing more predictable while preserving enterprise model-access controls.

The issue names no files, tests, or entry points. Start by locating the Auto model-selection settings and routing behavior in the app, then define how advisory preferences interact with availability, reliability, cost, and enterprise policy; done means users can configure task-aware preferences without bypassing model-access controls.

Written by the indexing model from the issue text.