Native Anthropic Messages chat model API

[maff]

What should we do?

Replace the LangChain4j bridge for the anthropic provider discriminator with a native implementation over the official com.anthropic:anthropic-java SDK. Direct backend only — Bedrock / Vertex / Foundry backends are out of scope for this issue.

Requirements:

• Streaming-first internal driver; public surface stays CompletableFuture<ChatResponse> plus the optional ChatStreamListener.
• Hardcoded ModelCapabilities matching today's behavior on this discriminator: text + image + document on user-message; text on tool-result; text on assistant-message.
• Token usage accounting populated from the provider response (cache + reasoning fields stay at zero in this phase).
• Stop-reason mapping from Anthropic finish reasons to the normalized StopReason enum.
• Transport / SDK / auth errors complete the future exceptionally with ConnectorException(ERROR_CODE_FAILED_MODEL_CALL, ...); model-side terminal errors (refusal, content filter, etc.) complete normally with stopReason=ERROR.
• HTTP transport via the shared JDK HttpClient utility so proxy config works on day one.

Why should we do it?

First native implementation; validates the SPI shape against a real vendor SDK before generalising to OpenAI, Google, Bedrock-Converse.