Add HPU support to the time-series MIA example (deferred from #425)

Follow-up to #425 (HPU support), which was deliberately scoped to exclude this.

The time-series MIA example (`examples/mia/time_series_mia/`) is not in place yet, so HPU enablement for it was left out of #425 rather than being written against a moving target. This issue tracks finishing it once the example itself is settled.

### What #425 already did touch

- `examples/mia/time_series_mia/data_handler.py`, `main.py`, `utils/model_preparation.py` and `main.ipynb` got the mechanical `get_device()` / `mark_step()` swap, so they are not on the old hardcoded-CUDA path.
- `leakpro/signals/utils/get_TS2Vec.py` explicitly falls back to CPU on HPU, because TS2Vec upstream has no HPU support:

```python
# TS2Vec upstream does not support HPU; fall back to CPU for this model.
```

So the example will run on an HPU box, but the representation-learning step silently runs on CPU.

### What still needs doing

1. Decide whether the TS2Vec CPU fallback is acceptable for the DTS attack, or whether the representation step needs an HPU-capable replacement. If it stays on CPU, that belongs in the docs, not just a log line — it's a performance cliff in the middle of an otherwise accelerated run.
2. Validate the example end to end on a Gaudi card once the example is in place, and record which `PT_HPU_LAZY_MODE` it was tested in.
3. Confirm `mark_step()` placement in the time-series training loops matches the pattern used elsewhere (after `optimizer.step()`).

### Not blocking #425

#425 should merge without this. Whoever picks up the time-series example should pick this up with it.