Migrate Mistral AI chat model to message parts

[dimitarproynov]

MistralAiChatModel now builds the assistant message of each choice from ordered message parts. Mistral returns the content as an ordered list of chunks: a thinking chunk becomes a ReasoningPart with the mistral_ai provider attribute and its signature, if any, as an opaque payload; a text chunk becomes a TextPart; and a reference or image chunk becomes an UnknownPart holding the chunk as JSON. The tool calls, which Mistral returns in a separate field, follow as ToolCallParts. Streamed parts are stamped with their index per choice, so the MessageAggregator merges each run of deltas into one part.

An assistant message is replayed as content chunks in part order. Mistral recommends sending the thinking of its reasoning models back, so a Mistral ReasoningPart goes back as a thinking chunk with its signature, while reasoning from another provider is not sent. A message without reasoning parts, such as one built by an earlier version, falls back to the thinking_content metadata, as OpenAiChatModel does with its reasoning metadata. User message parts are sent in order, so text and images can interleave.

MistralAiApi.ThinkChunk gains the signature field of the API, with a constructor that keeps the previous components. getText() now joins the text chunks without the line separator, which split the text around a citation. Streamed chunks always carry the response id, and THINKING_CONTENT_METADATA is deprecated in favor of getReasoning().

Update the Mistral AI reference documentation and the upgrade notes.

Build, unit and integration tests pass.