feat(llm, openai): Add GPT-5.6, max reasoning, prompt caching

[JeanMertz]

JP now recognizes the GPT-5.6 model family (Sol, Terra, and Luna variants, plus the gpt-5.6 alias resolving to Sol) and adds a new max reasoning effort level for models that support it. Models without max support keep degrading to xhigh, so existing configs using reasoning.effort: max continue to work everywhere but now get the deeper level on models that expose it.

For OpenAI, a new reasoning_mode parameter (standard or pro) lets users opt into the Responses API's pro reasoning mode on supported models; unsupported models fall back to standard with a warning instead of failing the request. Supporting models also get persisted reasoning: encrypted reasoning items from earlier turns are replayed with reasoning.context: all_turns so the model can factor prior reasoning into the next sample.

Requests to OpenAI also carry a stable prompt_cache_key derived from the conversation's creation timestamp, to improve cache-hit rates. Models that require explicit prompt caching (prompt_cache_options/prompt_cache_breakpoint) get breakpoints after the system prompt and after attachments, and caching can be disabled entirely via the assistant's cache policy, which also disables the explicit-cache opt-in for models that bill cache writes.