Write path serializes thinking as reasoning_content which some providers don't support

In #981, reasoning effort was enabled for Ollama, which also updated the read path to check both `reasoning` and `reasoning_content` fields. However, when `preserve_thinking = TRUE`, the multi-turn write path in `ProviderOpenAICompatible` serializes `ContentThinking` back as `reasoning_content`. Providers like vLLM only recognize `reasoning` (removed `reasoning_content` entirely in vllm-project/vllm#33402). The ecosystem is split on field naming: - OpenAI and OpenRouter recommend `reasoning` as primary (with `reasoning_content` as alias) - LM Studio uses only `reasoning_content` - vLLM uses only `reasoning` One option is a field on `ProviderOpenAICompatible` to configure which field name to use. See discussion in #981 for more context. Not currently biting anyone since `preserve_thinking` defaults to `FALSE` for Ollama, but worth addressing.