fix: keep mldev candidate parts separate

[he-yufeng]

Summary

Fixes #2507.

The Gemini Developer API can return multiple candidates with the later candidates' parts duplicated onto candidates[0].content.parts. The SDK then exposes response.text as the concatenation of every candidate, even though the helper says it returns the first candidate.

This normalizes only that exact mldev shape: when the tail of candidate 0's parts matches the parts from the sibling candidates, the duplicated tail is removed before the response model is built. Ordinary multi-part first-candidate responses are left alone.

Tests

python -m py_compile google\genai\models.py google\genai\tests\types\test_part_type.py
python -m pytest google\genai\tests\types\test_part_type.py::test_mldev_candidate_count_keeps_candidate_parts_separate google\genai\tests\types\test_part_type.py::test_two_candidates_text -q
python -m pytest google\genai\tests\types\test_part_type.py -q
python -m mypy google\genai\models.py --config-file pyproject.toml
git diff --check

[he-yufeng]

Rebased this branch onto current main and pushed the updated head ed07486.

Validation:

• python -m pytest google\genai\tests\types\test_part_type.py -q -k "mldev_candidate_count_keeps_candidate_parts_separate" -> passed, 1 passed
• python -m py_compile google\genai\models.py google\genai\tests\types\test_part_type.py -> passed
• git diff --check upstream/main..HEAD -> passed