feat(core): configure reasoning response extraction

[EAGzzyCSL]

Summary

• Adds adapter-level message extraction configuration for chat completion responses, including configurable default reasoning keys and custom extraction hooks.
• Moves reasoning-as-content fallback behind useReasoningAsContentFallback so it is enabled only for supported model families.
• Enables the fallback for Doubao, Qwen 3.x, GLM, Kimi, and Xiaomi Mimo while leaving OpenAI/GPT, Gemini, Auto-GLM, UI-TARS, and Qwen 2.5 unchanged.
• Keeps the default extractor conservative: it reads reasoning_content only by default.
• Configures Qwen 3.x to read reasoning_content first and reasoning second, covering DashScope and vLLM's OpenAI-compatible field migration: [RFC]: reasoning_content -> reasoning vllm-project/vllm#27755.
• Keeps Gemini on its custom extraction path only: Gemini reads thought parts and inline <thought> content, and ignores provider reasoning_content fields.
• Adds unit coverage for extraction configuration, Qwen/vLLM-style reasoning fields, Gemini custom extraction, fallback behavior, and the model-family whitelist.

Validation

./node_modules/.bin/vitest --run tests/unit-test/model-adapter/gemini.test.ts tests/unit-test/model-adapter/chat-completion.test.ts tests/unit-test/service-caller-reasoning-fallback.test.ts tests/unit-test/model-adapter.test.ts

Latest result: 4 test files, 42 tests passed.

Note: pnpm --filter @midscene/core test ... was attempted earlier, but pnpm dependency-state checks triggered install/build-script approval flow before tests ran, so the focused Vitest command above was used for validation.