Allow user-defined rules for Auto model selection (task type → model/tier)

## Problem

"Auto" mode for model selection currently applies a single policy decided by the Copilot/VS Code team. Users have no way to influence which model or reasoning tier gets picked for a given task, even though we often know in advance what a task needs.

The result is a forced tradeoff: either accept Auto's opaque choices, or opt out of Auto entirely and hand-pick a model every session — which defeats the cost/convenience benefit Auto is supposed to provide.

## Request

Let users define their own routing rules that Auto respects, instead of a fixed built-in policy. For example, rules based on:

- **Task type / intent** (e.g. "quick lookup," "refactor," "multi-file migration," "test writing")
- **Reasoning effort** (high / medium / low)
- **Exact model** to use for a given rule, not just a tier

Concretely, this could look like a config block (settings.json or a dedicated Auto-mode settings UI) where a user maps conditions to a model + effort level, and Auto uses those mappings instead of (or as an override to) its default heuristic.

## Why this matters

- Teams have differing cost constraints and already know which tasks warrant a stronger (and pricier) model vs. a cheap one.
- Right now "Auto" only reflects what the harness maintainers decided Auto should be — not what any individual team's workflow or budget requires.
- This would let cost-conscious teams keep using Auto (for convenience) while still getting predictable, self-chosen behavior instead of a black box.