fix: 修复 B 站下载与任务状态文件异常

[iorireal]

Summary

This is a focused replacement for #332. It keeps the PR scoped to the user-facing backend failures from #330 and removes the unrelated agent/Claude docs, AI workflow files, frontend lint cleanup, Mermaid work, and large lockfile churn from the previous PR.

Changes

• Upgrade yt-dlp from 2025.3.31 to 2026.3.17 to pick up newer Bilibili extractor/Wbi handling.
• Route Bilibili and YouTube downloader logs through the project logger via get_logger(__name__).
• Apply shared Bilibili yt-dlp options for audio, video, and subtitle extraction: browser-like headers, extractor retries, and cookiefile support.
• Keep CookieConfigManager support and add a cookies.txt fallback lookup path for local/backend/Docker usage.
• Add skip_download support to BilibiliDownloader.download() so subtitle-first flows can extract metadata without downloading media.
• Harden task status polling against empty or corrupt status JSON files, including the case where the final result file already exists.
• Simplify status writes to avoid the Windows temp-file replace race that can leave an empty status file.
• Ignore local AI agent files and local cookies.txt files so future PRs do not accidentally include them.
• Add regression tests for empty/corrupt status files and Bilibili skip_download interface compatibility.

Fixes #330.

Validation

• python -m py_compile backend/app/downloaders/bilibili_downloader.py backend/app/downloaders/youtube_downloader.py backend/app/routers/note.py backend/app/services/note.py backend/tests/test_note_status_resilience.py
• cd backend && python -m pytest tests/test_note_status_resilience.py tests/test_note_helper.py -q -> 5 passed
• cd backend && python -m pytest tests/ -q -> 14 passed, 1 failed in tests/test_task_serial_executor.py::TestTaskSerialExecutor::test_executor_runs_tasks_one_by_one; this is the existing origin/master baseline mismatch where SerialTaskExecutor is currently an alias for ConcurrentTaskExecutor, not introduced by this PR.