fix: dataset code block splitting

[YYH211]

What

• Prevent fenced code blocks from swallowing large preceding text during dataset chunking.
• Cap code block preservation by chunk size instead of model max context.
• Add regression coverage for long text followed by markdown image code blocks.

Tests

• pnpm exec prettier --check packages/global/common/string/textSplitter.ts packages/global/test/common/string/textSplitter.test.ts
• pnpm exec eslint --fix packages/global/common/string/textSplitter.ts packages/global/test/common/string/textSplitter.test.ts
• pnpm --filter @fastgpt/global test -- common/string/textSplitter.test.ts
• pnpm --filter @fastgpt/app build:workers

[github-actions]

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	13.96%	1140 / 8161
🔵	Statements	13.95%	1195 / 8564
🔵	Functions	12.47%	245 / 1964
🔵	Branches	11.94%	536 / 4489

File Coverage

No changed files found.

Generated in workflow #454 for commit 8b5b89f by the Vitest Coverage Report Action

[github-actions]

✅ Admin Preview Image Ready!

ghcr.io/labring/fastgpt-pr:admin_8b5b89f4f4bd05405f822248124a8b7092e870b5

🕒 Time: 2026-05-27 14:49:11 (UTC+8)

[github-actions]

✅ Build Successful - Preview fastgpt Image for this PR:

ghcr.io/labring/fastgpt-pr:fastgpt_8b5b89f4f4bd05405f822248124a8b7092e870b5

🕒 Time: 2026-05-27 14:49:50 (UTC+8)