Tracing temp directories are never cleaned up when keywords are called from Python outside a Robot Framework run

Nobody has claimed this yet.

Tracing writes its working files to `browser/traces/temp//` (`Browser/keywords/playwright_state.py:911`). Those are cleaned up from the `_start_suite` listener hook (`Browser/browser.py:838-848`), which only fires during a Robot Framework run. When keywords are called directly from Python the hook never runs, so the temp directories accumulate and are never removed.

### Reproduce

```python
from Browser import Browser

browser = Browser()
browser.new_browser("webkit", headless=True)
browser.new_context(tracing=True)
browser.new_page("https://robotframework-browser.org")
browser.close_context()
browser.close_browser("ALL")
```

Observed afterwards, left in place:

```
browser/traces/temp//
.trace
.network
resources/*.jpeg
```

The trace zip itself is produced correctly by `close_context()`; it is only the temp working directory that is left behind, once per traced context.

Note that with the default `outputdir` these land inside the installed package rather than the caller's working directory — that part is a separate issue about the relative default `outputdir`, and it is filed on its own. This report is about the cleanup never happening at all, which remains true even when `outputdir` is absolute.

### Suggested fix

Remove the temp directory when the trace is flushed in `close_context()`, rather than relying on a listener hook that only exists during a Robot Framework run.

### Environment

Browser 20.3.0, Robot Framework 7.4.1, Python 3.14.7, macOS, webkit.

Start in Browser/keywords/playwright_state.py around line 911 and follow close_context() through its trace flush. Compare that path with the cleanup in Browser/browser.py:838-848, which only runs from the _start_suite listener hook. Reproduce the Python-only example and confirm that closing a traced context leaves no directory under browser/traces/temp//.

Written by the indexing model from the issue text.