feat(custom):新增 screenshot_on_fail.py节点级截图sink - #14

Please address the comments from this code review:

## Individual Comments

### Comment 1
<location path="Storage/customs/BQ/screenshot-on-fail/screenshot_on_fail.py" line_range="130" />
<code_context>
+        if noti_type in (NotificationType.Succeeded, NotificationType.Failed):
</code_context>
<issue_to_address>
**issue (bug_risk):** The sink saves a screenshot for every notification type, but only retrieves node-specific recognition data for `Succeeded` and `Failed`. When the callback receives `Starting`, it falls back to `cached_image` and labels the resulting file `failed`, producing a misleading failure screenshot for an incomplete recognition event.

**Triggers:** When `on_node_pipeline_node` emits its normal `Starting` notification.

**Suggested fix:** Return unless `noti_type` is `NotificationType.Succeeded` or `NotificationType.Failed`, or handle `Starting` with its own status and image semantics.

```suggestion
        if noti_type not in (NotificationType.Succeeded, NotificationType.Failed):
            return

        node_name = detail.name or "unknown"
```
</issue_to_address>

### Comment 2
<location path="Storage/customs/BQ/screenshot-on-fail/screenshot_on_fail.py" line_range="158-166" />
<code_context>
+        filename = f"{timestamp}_{node_name}_{reco_id}_{status}.jpg"
+        filepath = dirpath / filename
+
+        _save_image(img, filepath)
+        _cleanup_old_screenshots()
\ No newline at end of file
</code_context>
<issue_to_address>
**issue (bug_risk):** Failures from image encoding, file creation, and cleanup are not handled at the callback boundary. `cv2.imencode`, `open`, `mkdir`, `unlink`, or `_cleanup_old_screenshots` can raise, causing the context sink callback to fail instead of merely reporting that a debug screenshot could not be saved.

**Triggers:** When the debug directory is unwritable, the disk is full, a screenshot has an unsupported dtype/shape, or cleanup races with another filesystem change.

**Suggested fix:** Wrap saving and cleanup in exception handling, log the failure, and ensure screenshot diagnostics cannot interrupt task execution.

```suggestion
        try:
            dirpath = Path(_SCREENSHOT_DIR)
            dirpath.mkdir(parents=True, exist_ok=True)
            timestamp = datetime.now().strftime("%Y.%m.%d-%H.%M.%S.%f")[:-3]
            status = "success" if noti_type == NotificationType.Succeeded else "failed"
            filename = f"{timestamp}_{node_name}_{reco_id}_{status}.jpg"
            filepath = dirpath / filename

            _save_image(img, filepath)
            _cleanup_old_screenshots()
        except Exception:
            _log.exception("保存调试截图失败")
```
</issue_to_address>

### Comment 3
<location path="Storage/customs/BQ/screenshot-on-fail/screenshot_on_fail.py" line_range="70" />
<code_context>
+            f.write(encoded.tobytes())
+        return True
+
+    return _save_bmp(img, filepath.with_suffix(".bmp"))
+
+
</code_context>
<issue_to_address>
**nitpick (bug_risk):** The fallback writes to `filepath.with_suffix(".bmp")`, while the caller continues to use and document the original `.jpg` path. The documented output pattern therefore does not match the actual fallback output, and any caller or tooling expecting the returned `.jpg` filename has no way to discover the generated BMP file.

**Triggers:** When cv2 is unavailable.

**Suggested fix:** Return the actual output path from `_save_image`, construct the extension before writing, and keep the README/docstring filename pattern consistent with the fallback behavior.
</issue_to_address>