[Bugfix] Preserve valid HTML attributes in CCOCR table evaluation

[Sun-Jiatuhao]

Summary

This PR fixes abnormally low CCOCR table parsing scores for model outputs containing valid HTML table attributes or Markdown code fences.

Some otherwise correct table outputs, such as tables containing border, colspan, or rowspan attributes, could receive unexpectedly low or zero scores during evaluation.

Changes

• Preserve valid HTML tag and attribute syntax during preprocessing.
• Extract HTML tables from Markdown fences and surrounding response text.
• Normalize whitespace in text nodes without modifying HTML syntax.
• Treat th elements as table cells during TEDS evaluation.
• Handle malformed or missing tables safely.
• Add regression tests for the affected output patterns.

Tests

python -m unittest -v tests/test_ccocr_table_evaluator.py

All 4 tests pass.