Allow per-prompt model and provider configuration

[Neokil]

Problem

Every prompt in a workflow uses the same provider and model, set globally in config. There's no way to route different steps to different models — e.g. a powerful reasoning model for the investigation phase, a coding-optimized model for implementation, and a cheap/fast model for mechanical tasks like generating commit messages. This makes it impossible to optimize cost and quality per step.

Solution

1. Add optional provider and model fields to the workflow state config (alongside the existing prompt field).
2. When executing a state, use the state-level provider/model if set, falling back to the global config default.
3. Validate at config load time that any referenced provider names exist in the providers map.
4. Document the per-state overrides in the workflow config example file.