fix(scicode): read assistant text blocks

[he-yufeng]

Summary

• use ChatMessage.text when collecting SciCode assistant solutions
• keep the existing markdown code-fence extraction behavior unchanged
• add regression coverage for assistant messages whose content is stored as text blocks

Fixes #1380.

To verify

• python -m py_compile evalscope\benchmarks\scicode\util.py tests\benchmark\test_scicode_util.py
• python -m pytest tests\benchmark\test_scicode_util.py -q
• python -m flake8 evalscope\benchmarks\scicode\util.py tests\benchmark\test_scicode_util.py
• python -m yapf --diff evalscope\benchmarks\scicode\util.py tests\benchmark\test_scicode_util.py
• git diff --check

[gemini-code-assist]

Code Review

This pull request simplifies the code extraction logic in evalscope/benchmarks/scicode/util.py by accessing message.text directly instead of casting message.content to a string. It also adds a unit test in tests/benchmark/test_scicode_util.py to verify that get_generated_code correctly handles assistant messages with text content blocks. There are no review comments to address, and the changes appear correct and complete.

[Yunnglin]

LGTM. Fixed a minor yapf formatting issue and pushed the fix.

[Yunnglin]

I pushed a small formatting fix (yapf) to your branch — the list comprehension in util.py and the test file needed minor style adjustments to pass make lint. Everything else looks good!