fix: stop restarting exhausted OpenAI rate-limit retries - #1204

The reporter runs PDF2zh 1.9.11 on Windows 11/Python 3.12.11 and sees repeated OpenAI RateLimitError warnings while the GUI appears stuck; Google translation works. The supplied logs name translator.py and show exponential retry delays with a 100-attempt budget. Current OpenAITranslator.do_translate still has that budget, but TranslateConverter.receive_layout wraps its paragraph worker in an unlimited retry decorator, restarting exhausted translation attempts forever. The only reply attributes the rejection to the API account's limits; the response payload is absent, so neither insufficient quota nor a transient request limit is established.

Testing

OpenAI succeeds immediately in streaming and non-streaming modes: text and normal request options remain unchanged.
A genuine mocked openai.RateLimitError followed by success retries and returns the translation; construct errors with a synthetic HTTP 429 response and no credentials or external requests.
Persistent RateLimitError exhausts exactly the configured 100 application-level attempts with sleeping disabled and raises the original final exception, not tenacity.RetryError. Mock the SDK call directly so SDK-internal retries do not distort the expected count.
Feed a minimal nonempty text layout into real TranslateConverter.receive_layout with one worker and a real OpenAITranslator whose mocked client always fails: the combined path exits after the same 100 calls, never starts call 101, and does not emit translated output. Use a fail-fast sentinel on call 101 so the pre-fix regression cannot hang even under the unlimited outer decorator; do not mock receive_layout itself.
A non-rate-limit ordinary exception followed by success retains the converter's existing retry behavior; blank text/formula-only paragraphs remain skipped.
Run the focused translator and converter suites with existing installed dependencies (pytest test/test_translator.py test/test_converter.py), then the project's unit suite and Black check for changed files in the implementation lane. No live API quota, PDF downloads, or model downloads are required for the new tests.

Summary

Set reraise=True on the existing OpenAITranslator.do_translate Tenacity decorator so exhausting its existing 100-attempt budget propagates the original openai.RateLimitError instead of a RetryError wrapper. In pdf2zh/converter.py, import the exception and configure the existing paragraph worker retry predicate to exclude RateLimitError while retaining retries for other ordinary Exceptions; this caller change is essential to prevent resetting the translator's budget. Preserve the existing exponential backoff, streaming behavior, subclass inheritance, and attempt count, and introduce no new retry abstraction or test-only seam.

Fixes #1147