fix: clean up Playwright process on browser startup failure

[Battleplus]

这次改了什么

When Playwright starts successfully but Chromium launch fails (e.g. missing browser binary, sandbox issues), the node cli.js run-driver subprocess was leaked. Each failed attempt left one live child process, growing the process count linearly with retries.

Changes:

1. browser_manager.py: Extract browser launch logic into _launch_browser(), wrap the call in start() with try/except that calls self.playwright.stop() on failure.

2. async_webcrawler.py: Wrap crawler_strategy.__aenter__() in start() with try/except that calls __aexit__() on failure.

Both layers ensure partial resources are cleaned up before the exception propagates.

怎么验证的

• Python syntax check passes for both files
• Manual verification: AsyncWebCrawler.__aenter__() now properly cleans up when browser launch fails
• No live playwright driver processes remain after failed start attempts

风险

Low. Only adds cleanup on the exception path. Happy path is unchanged.

Closes #2155

[Battleplus]

Follow-up audit: startup rollback now catches BaseException, not only Exception, so task cancellation during Playwright/browser startup cannot leak the driver. Added unit coverage for ordinary launch failure, asyncio cancellation at BrowserManager startup, and cancellation while entering AsyncWebCrawler. Validation: 3 passed. Commit: 15a14c9.

[Battleplus]

Closing as a duplicate after the full audit. Earlier PR #2160 wraps the complete BrowserManager startup sequence in BaseException-safe rollback and includes a real Playwright child-process leak regression. This PR was additionally hardened for cancellation and its unit tests passed, but #2160 is earlier and covers more startup branches, so review should stay there.