When a user switches to a model that doesn't support thinking (e.g.,
Gemini 1.5 Flash) while a thinking level is still set from a previous
model, the app crashes with an uncaught exception from
requireSupportedEffort().
Add model.reasoning guards in mapOptionsForApi() for all provider paths
(Anthropic, Google, Google Gemini CLI, Google Vertex, OpenAI) so that
thinking is gracefully disabled instead of throwing. The Bedrock path
already had this guard.
