fix: add provider-safe structured output mode

[radu-mocanu]

Summary

• add opt-in method="auto" structured output for the normalized chat client
• use JSON mode for Anthropic models and sequential function calling for other providers
• preserve every explicit method and the existing function_calling default
• make include_raw=True work with standard LangChain message-list inputs

Why

The normalized API unifies transport, but providers still differ in structured-output capabilities: Gemini rejects the OpenAI response formats used by JSON mode and JSON schema, while Claude Haiku can return an incomplete schema through function calling even when parallel calls are disabled. Centralizing the capability choice in this client keeps Health Score and other consumers provider-independent.

Validation

• uv run ruff check
• uv run ruff format --check
• uv run pyright .
• normalized client unit tests: 17 passed, 1 skipped, 1 expected xpass
• live Alpha validation: 6/6 passed with Haiku, Gemini 2.5 Flash, and GPT-4o across both agentsplaygrounddataresidency and agentsplaygroundfallback