Add `system_instruction` field to `LLMSettings` so it is runtime-updatable via settings.
For Google (GoogleLLMService, GoogleVertexLLMService), deprecate the init-time arg since it was already shipped. For Anthropic, AWS Bedrock, and OpenAI, remove the init-time arg entirely since it was never shipped.
Add system instruction prepend logic to `build_chat_completion_params` overrides in Cerebras, SambaNova, Fireworks, Mistral, and Perplexity, which build params from scratch rather than calling `super()`.
Still need to handle realtime services (OpenAI Realtime, Grok Realtime, Gemini Live).
Prior to this change, after the model generated an image the conversation would not be able to progress. It would stall out because we were never storing the image in context, so the model would never realize it already did the work of generating an image. We didn't run into issues with Gemini 2.5 Flash Image, because that model always followed up an image with a text message.
