make long-file CUDA OOM recoverable per file

- replace permanent _CUDA_BROKEN flag with CudaOOMError and retry helper
- retry WhisperX with configured settings, then batch=1 without VAD, then once more
- keep the loaded model cached between retries instead of reloading it
- actually pass vad_filter through transcribe() and use it in aligned ASR
- skip only the OOM file in sequential, staged and LRC pipelines
- mark OOM files with oom_skipped / lrc+oom for the summary
- fix staged vocal separation writing input_dir/vocals.wav when convert is off
- route original inputs and resumed files through safe per-file separation
- validate ffmpeg before starting the LRC task

gui/main_window.py

@@ -692,6 +692,14 @@ class MainWindow(QMainWindow):
                                     f"Команда '{cfg.separate.unblend_path}' не найдена в PATH.")
                 return
 
+        if task == "lrc" and not shutil.which(cfg.convert.ffmpeg_path):
+            QMessageBox.warning(
+                self, "ffmpeg не найден",
+                f"'{cfg.convert.ffmpeg_path}' не найден в PATH.\n"
+                "LRC-этапу нужен ffmpeg для подготовки аудио."
+            )
+            return
+
         if task == "lrc" and cfg.transcribe.engine != "whisperx":
             QMessageBox.warning(
                 self, "LRC требует WhisperX",