Mistral AI : fix responses from reasoning models

[dskarzh]

Issue:

Mistral AI reasoning models (Magistral) return content as an array of content objects, whereas non-reasoning models return it as a string. LangChain4j only handled string responses, causing deserialization failures with reasoning models.

Changes:

• Add MistralAiThinkingContent and MistralAiReferenceContent classes for new content types
• Add polymorphic deserialization support to MistralAiMessageContent via @JsonTypeInfo/@JsonSubTypes
• Update MistralAiChatMessageBuilder.content() to handle both string and array (reasoning models) formats
• Update MistralAiChatMessage.asText() to extract text content from multi-part responses