feat: add per-model reasoning effort setting

[dongyou03]

Summary

• add an optional Reasoning effort selector to each editable LLM model
• persist the selection as the existing model-level reasoning_effort field
• make Auto remove the field so provider defaults remain unchanged
• offer conservative provider/model-specific choices instead of one misleading universal enum
• localize the new control and guidance in English and Chinese

Why

DeepTutor's runtime already resolves reasoning_effort from the active catalog model and maps it into each provider's request shape, but the Web model editor did not expose the field. Users therefore had to edit model_catalog.json manually.

This is the model-default layer discussed in #641; a future per-conversation override can take precedence over it.

Safety

• Gemini 3 and Gemini 2.5 Pro offer minimal, not the invalid none value reported in Gemini 3 models fail with 400 INVALID_ARGUMENT — reasoning_effort="none" is only valid for Gemini 2.5 non-Pro #734
• Gemini 2.5 non-Pro models may explicitly select none
• OpenAI, Anthropic, and binary thinking providers receive conservative family-specific choices
• unknown model families remain hidden unless an existing manual override needs to be displayed or cleared
• Auto deletes reasoning_effort rather than persisting an empty string

Validation

• pnpm --dir web run test:node — 368 passed
• pnpm --dir web exec tsc --noEmit
• targeted ESLint — 0 errors (6 pre-existing placeholder warnings in ServiceConfigEditor.tsx)
• pnpm --dir web run i18n:check — locale parity OK
• pnpm --dir web build
• ./.venv/bin/pytest tests/services/llm/test_reasoning_params.py — 17 passed
• ./.venv/bin/pytest tests/services/config/test_provider_runtime.py -k reasoning_effort — 1 passed
• Impeccable detector — no findings
• browser verification on /settings/llm:
  • OpenAI gpt-5.2 shows Auto, Minimal, Low, Medium, High, and Extra high
  • Gemini gemini-3.6-flash shows Auto, Minimal, Low, Medium, and High, with no none
  • selecting High and returning to Auto works
  • Chinese copy renders correctly and the console has zero errors

Fixes #735
Related: #641, #734