feat(models): add Google Cloud Translation as a LID model

[malteos]

Closes #11.

Why

Adds GoogleTranslate, so the commercial detector can be compared against the open models already on the leaderboard.

Builds on #17: the wrapper returns the API's confidence as the prediction score.

v2, not v3

The issue links the v3 codelab, but v3 cannot be used with an API key. detectLanguage refuses it outright on both transports:

401 API keys are not supported by this API. Expected OAuth2 access token
or other authentication credentials that assert a principal.
[reason: "CREDENTIALS_MISSING"]

It wants OAuth2 / Application Default Credentials plus a project id in the parent path. v2 accepts an API key and needs no project id.

v2 is also the better fit on the merits. It detects a whole list of texts per request, where v3 takes a single content string per call and batchTranslateText covers translation only. That turns a 373K-sample benchmark from 373K HTTP calls into a few thousand.

The client library ignores client_options={"api_key": ...} and falls through to ADC, so the key is passed as google.auth.api_key.Credentials, which attaches the x-goog-api-key header.

Credentials

GOOGLE_TRANSLATE_API_KEY is read from the process environment and nothing else. The model does not load .env files; setting the variable is the caller's job. A missing key fails at load() with a message naming it.

Cost

Detection is billed per character at the same rate as translation, $20 per million characters, with the first 500K each month free. Measured on the cleaned text that is actually sent, since predict() applies the OpenLID normer first:

dataset	samples	billable chars	cost
commonlid_nano	1,507	305,431	~$6
commonlid	373,230	76,713,413	~$1,534

The nano run fits inside the monthly free tier. No evaluation has been run, and none should be without agreeing the spend first. Re-runs are free because the evaluator's PredictionCache covers them.

Implementation notes

• Each batch is re-split to stay under the per-request segment and size caps, and the pieces go out concurrently. Raise --batch-size for larger requests.
• Retries 429 and transient errors with bounded backoff. A dead endpoint or a bad key raises rather than being silently scored as abstention.
• Blank text short-circuits before the call: a rejected request is still billed.
• Detected BCP-47 codes are cut at the first - before the existing ISO 639-3 conformance pass, so zh-CN becomes zho. Google's legacy codes are already handled by the deprecation table, so iw becomes heb.
• discover_supported_languages() asks the API, giving 187 conformed codes from 195 raw, with nothing dropped by the conformance pass.
• New optional extra commonlid[google-translate], named after the service rather than the vendor so a future Google model gets its own. Mirrored into dev so the tests exercise the real client, and added to all. New make install-google-translate.

Testing

make check passes: 277 tests, 95.8% coverage, ruff and mypy strict clean.

The model tests never touch the network; they swap in a fake client or monkeypatch the library out of sys.modules. They cover the missing-extra and missing-key errors, code conformance including zh-CN and a confidence surviving an unmappable code, blank-text short-circuiting with zero API calls, chunking by both segment and character budget, order preservation across the thread pool, and the retry and give-up paths.

Verified against the live API. A three-sample evaluator run, costing a few hundred characters, produced:

{"correct": true, "gold": "lav", "idx": 2, "pred": "lav", "score": 0.6271151304244995, ...}