feat: add Serply as optional Google Scholar search backend

[googio]

Summary / 概述

Adds serply as an optional keyed search backend: Serply's Google Scholar API,
wired into search.backends exactly like serpbase.

The motivation is corpus coverage for the novelty audit. alphaxiv is the only
papers backend today and it is arXiv-only, so prior art that exists solely in
venue proceedings or a journal is invisible to it. Scholar indexes that
published record, so backends: [alphaxiv, serply] merges and de-duplicates
into a wider prior-art view without changing anything about how the audit runs.

A live query confirms the gap is real rather than theoretical: the top hit for
"iterative hypothesis generation scientific discovery" is an ACL Anthology
paper with no arXiv mirror in the result set.

Linked issue / 关联 issue

Closes #66

Type of change / 改动类型

• 🐞 Bug fix / 缺陷修复 (fix)
• ✨ Feature / 新功能 (feat)
• 📝 Docs / 文档 (docs)
• ♻️ Refactor / 重构 (refactor)
• ✅ Tests / 测试 (test)
• 🔧 Chore / build / CI (chore)

How was this tested? / 如何验证？

Baseline numbers below are from upstream/main in the same virtualenv, so the
comparison is apples to apples.

$ python -m pytest tests/test_search_backends.py
36 passed in 0.20s                      (29 on main, so +7 new)

$ python -m pytest tests/
504 passed, 3 skipped in 7.20s          (main: 497 passed, 3 skipped)

$ python -m ruff check src/ tests/
All checks passed!

$ python -m mypy src/
Found 242 errors in 36 files            (main: 242 -- none introduced)

The unit tests mock requests.get; no network is touched. Separately I ran the
backend against the real API through resolve_backend_names /
build_search_backends to check the response contract:

HTTP 200 | len(articles) = 3 | len(results) = 0
{
  "title": "Hypothesis generation with large language models",
  "link": "https://aclanthology.org/2024.nlp4science-1.10/",
  "description": "Y Zhou, H Liu, T Srivastava, H Mei… - … on NLP for Science  …, 2024 - aclanthology.org"
}

resolve_backend_names(backends=[alphaxiv, serply], key set) -> ['alphaxiv', 'serply']
resolve_backend_names(backends=[alphaxiv, serply], no key)  -> ['alphaxiv']

Two things that response settles, and both are pinned by a test:

• Scholar results arrive in articles, and the envelope also carries an
  always-empty results list. Keying the wrong array returns zero hits with no
  error, so the array name is passed explicitly rather than guessed.
• description is the Scholar byline ("Authors - Venue, Year"), not an
  abstract. It maps to snippets because that is the field the contract has,
  but the docs say plainly what it contains so nobody expects abstract text.

Size / 体量

7 files, +195/-22. Slightly above the merged serpbase backend (#65), which
was 5 files, +133/-21, and the extra is accounted for:

• the same 5 files that PR touched, plus
• docs/search.zh.md and docs/configuration.zh.md (+14/-5 combined), the
  Chinese twins, because [Feature]: add Serply Google Scholar as an optional papers search backend #66 said the docs would be bilingual.

Two deliberate omissions, both easy to add if you would rather have them:

• I did not add serpbase to the .zh docs. It is missing there, but it is
  not my change to make in this PR.
• I did not touch README.md / README.zh-CN.md. Neither names any keyed
  backend today; both document only the legacy builtin_backend: none | alphaxiv path, so serply has nowhere natural to go. Same for
  examples/research_config.example.yaml, which has no search.backends
  block. serpbase was left out of all three as well.

Happy to split the .zh docs into a follow-up if you prefer this PR at
precedent size.

Notes

• Serply support is optional, and behavior is unchanged when SERPLY_API_KEY
  is unset: a keyed backend with no key is dropped in resolve_backend_names,
  so backends: [alphaxiv, serply] degrades to [alphaxiv] silently, the same
  as serper and serpbase.
• Follows SerpBaseBackend in src/core/tools/web/backends.py line for line:
  same _SyncBackend base, same _HTTP_TIMEOUT, same raise_for_status, same
  {url, title, snippets} output, same truncation to max_results. It is a
  GET with an X-Api-Key header where SerpBase is a POST.
• Key resolution mirrors the existing helpers: serply_api_key on
  SearchConfig, falling back to SERPLY_API_KEY via _WebSearchEnv.
• Requests identify the caller with User-Agent: Arbor.
• Tests in tests/test_search_backends.py; network mocked.
• Docs in docs/search.md, docs/configuration.md and their .zh twins.

Checklist / 检查清单

• PR title follows Conventional Commits (e.g. fix(config): ...). / PR 标题遵循 Conventional Commits 规范。
• The change is focused and reasonably small. / 改动聚焦且体量合理。
• I ran the relevant tests locally (python tests/test_*.py) and they pass. / 我已在本地运行相关测试并通过。
• Docs / examples updated if behavior or config changed. / 若行为或配置有变更，已同步更新文档/示例（含 README.zh-CN.md）。
  • Both READMEs and the example config are the exception, deliberately: see
    Size for why serply has no place in them today.
• No secrets, API keys, or tokens are included in the diff. / diff 中不包含任何密钥、API key 或 token。

Disclosure: I work with Serply. Happy to adjust scope, naming, or drop this
entirely if it isn't a direction you want for the project.