Make handling of Reasoning configuration and output across providers consistent

Goose has a shared thinking_effort setting and a reasoning model capability, but their meaning and behavior differ depending on the provider, request format, model name, and code path.
Reasoning support is currently determined through a mix of:

* Explicit model metadata
* Canonical model metadata
* Hard-coded model-name checks
* Provider-specific exception lists
* Request parameters
* Legacy environment variables

The resulting behavior is difficult to predict, especially for custom providers, aliases, newly released models, and models exposed through a provider other than their original vendor.

We should come up with a way to make this work consistent across all providers rather than stacking ad hoc solutions on top of each other