support preserving full reasoning history for third party llms

some models like kimi k3 always require passing back `reasoning_content`s for *optimal performance*
ref: https://platform.kimi.ai/docs/guide/use-thinking-models#q1-why-should-i-keep-reasoning_content
note this means always preserve every `reasoning_content` in the session, not just inside current toolcalling loop

currently it seems neither chatcompletion byok models w/ chatLanguageModels.json nor the language models
extension api provides full access to reasoning history (however some internal copilot models seem able to do