[Python] Reset BigQuery retries for each batch

[bvolpato]

BigQuery streaming inserts share one retry iterator across an entire bundle and advance it even for successful requests. With max_retries=1, a successful batch exhausts the retry budget for the next batch: a transient failure then goes straight to the failed-row output. Prior successes also inflate the initial backoff for later batches; after 20 successful batches it reaches 750–1500 seconds.

Create the retry iterator per batch, matching the documented per-group retry limit. Partial-failure retries still preserve the failed rows and their insert IDs. Tests cover consecutive batches, different destinations, a previously exhausted batch, initial backoff, and zero retries.

Testing

From sdks/python, with the SDK and GCP test dependencies installed:

python -m pytest apache_beam/io/gcp/bigquery_test.py -q \
  -k 'successful_batches or exhausted_batch or zero_retries'

All 5 new cases pass. Against the upstream implementation, 4 fail and the zero-retry compatibility case passes.

All 50 tests in the three affected streaming test classes also pass on Python 3.12. Credential discovery was stubbed only in the test process; the tests' existing request/client mocks remain in use:

from unittest import mock
import pytest
from google.auth.credentials import AnonymousCredentials

with mock.patch('google.auth.default',
                return_value=(AnonymousCredentials(), 'test-project')):
    raise SystemExit(pytest.main([
        '-q', '--timeout=30',
        'apache_beam/io/gcp/bigquery_test.py::BigQueryStreamingInsertTransformTests',
        'apache_beam/io/gcp/bigquery_test.py::BigQueryStreamingInsertsErrorHandling',
        'apache_beam/io/gcp/bigquery_test.py::PipelineBasedStreamingInsertTest',
    ]))

The full module encountered an unrelated credential-lock failure in an existing read test, also reproduced on upstream. YAPF 0.43.0, Ruff 0.15.22, and git diff --check pass for the changed files.

Downsides

The retry allowance now applies independently to each batch, so several failing batches can make more total attempts than the previous accidental shared limit. The configured per-batch limit remains unchanged.

---

• Describe the bug and include reproducible regression tests.
• Update CHANGES.md with the behavior change.
• Apache Individual Contributor License Agreement, if required for this contribution.