refactor(llm): type provider tool choices

[Jordan-Jarvis]

Summary

• replace the internal str | dict tool-choice union with a canonical typed value
• serialize provider-specific tool-choice wire shapes only at provider boundaries
• make named-tool validation explicit and remove runtime payload-shape guessing
• preserve provider retry configuration and response-usage tracing from current upstream main

Dependency

This is intentionally stacked on #2734 because it types the named-tool behavior introduced there. Review the final refactor commit 7a71d8c11; after #2734 lands, the base will contain the prerequisite behavior and this PR will show only the typed cutover.

Validation

• Ruff check passed on all 17 changed files
• Ruff format check passed on all 17 changed files
• 41 focused provider/Reflect tests passed sequentially
• engine compileall passed

A broader local file-level run reached 82 passing tests; four unrelated integration/setup cases failed because the local environment had no .env API key and two parallel workers timed out waiting on the shared embeddings fixture lock. The focused deterministic suite above is the authoritative change validation.

---

Replacement for #2737. The original PR became unreopenable after its submitting fork was deleted. This PR restores the exact reviewed head commit and preserves the original implementation without modification.

[Jordan-Jarvis]

Ready for maintainer merge. This is the exact restored replacement for unreopenable #2737. Head 54f14f3 is already merged with current upstream main e07ca1b; the merge tree was independently reconstructed and verified byte-for-byte. All upstream checks are green. Please merge this one first because #2844 builds on the typed tool-choice contract.

[Jordan-Jarvis]

cc @nicoloboschi — all checks are green; this is first in the merge order.

[nicoloboschi]

Approve. Behavior-preserving typed refactor: verified provider serialization maps 1:1 to the old string/dict logic across all providers, ty check passes, and the four tool-choice unit suites pass (27 tests). Non-blocking follow-ups: Anthropic provider still ignores tool_choice (pre-existing); a small direct unit test for LLMToolChoice invariants would be nice.