Remove unnecessary thought signatures from content sent to the Gemini models

Gemini models only look at the the latest thought signature. All thought signatures prior to the latest one can be removed from the LLM request content to reduce input tokens and API request size.