llm: mock Anthropic backend (test fixture)

[jfallows]

Summary

A mock Anthropic-shaped backend server (streaming + non-streaming /v1/messages) for use in cross-dialect integration tests and (later) examples/llm.proxy, mirroring the mock OpenAI backend from M2.

Scope

• Streaming responses: realistic named-typed-event sequences (message_start, content_block_start, content_block_delta, content_block_stop, message_delta, message_stop), multi-block content, input_json_delta tool streaming.
• Non-streaming responses: top-level response object including usage.
• Deterministic/scriptable enough to drive the round-trip and cross-dialect tests in this milestone.

Acceptance criteria

• Mock server usable from IT tests and a docker-compose-style example, mirroring the mock OpenAI backend's structure for consistency.
• Covers both streaming and non-streaming Anthropic response shapes, including multi-block and tool-call content.

Depends on

dialect: anthropic (this milestone, for response shapes to mirror), mock OpenAI backend (M2, as the sibling to mirror).