No way to tell if prediction is running — backend silently pegs CPU (~800%) with zero status feedback

When requesting predictions for several tasks at once, there's **no signal that anything is happening**: the Label Studio UI shows nothing, yet the ML backend process sits at ~800% CPU in the background for a long time. You can't tell whether it's working, stuck, or dead.

**Why:** `/predict` is fully synchronous with no progress/status endpoint. gunicorn threads (`THREADS=8` by default) run multiple `predict()` calls concurrently, and some examples (e.g. `ppocr`) reload the model on every request — so the process thrashes CPU with no visibility.

**Ask:**
- A lightweight status/progress signal (`queued / processing / done / failed` per task) so users can tell it's actually working.
- Concurrency control (serialize or bound inference) to avoid the runaway CPU.

Start at the /predict entry point and trace how synchronous predict() calls are handled under the default gunicorn THREADS=8 setting. Review the ppocr examples mentioned in the issue, then define how per-task queued, processing, done, and failed status should be exposed and how inference concurrency should be bounded.

Written by the indexing model from the issue text.