Windows smoke tests fail on Greek text when DejaVu font path is unavailable

### Environment

- Windows 10
- Python 3.11.9
- PyTorch 2.5.1+cu121
- CUDA available with NVIDIA GeForce GTX 960M
- `glossapi_rs_cleaner` built and installed successfully

### Reproduction

From the `development` branch:

```powershell
python -m pytest tests/test_pipeline_smoke.py -rs
```

The test suite starts correctly and CUDA is detected, but two smoke tests fail while generating PDFs containing Greek text.

### Observed behaviour

The failures originate in `fpdf` with errors such as:

```text
FPDFUnicodeEncodingException: Character "Κ" ... is outside the range of characters supported by the font used: "helvetica"
```

The test helper currently uses:

```python
_FONT_PATH = Path("/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf")
```

On Windows that Linux-specific path does not exist, so `_write_pdf()` falls back to Helvetica. Helvetica cannot encode the Greek strings used by the smoke tests, which causes the failures before the actual pipeline assertions are reached.

### Expected behaviour

The smoke tests should be able to generate their Unicode test PDFs on supported development platforms without depending on a Linux-only font path.

### Possible direction

A platform-independent Unicode font lookup for the tests, or a small bundled test font, would avoid the fallback to Helvetica and make the smoke tests portable across Windows and Linux.

I found this while setting up GlossAPI on Windows as a new contributor. I am happy to work on a small fix if this approach makes sense to the maintainers.

Start with tests/test_pipeline_smoke.py and its _write_pdf() helper, then run `python -m pytest tests/test_pipeline_smoke.py -rs` on Windows. Check how the DejaVuSans path is selected and how the helper falls back to Helvetica. Done means the Greek-text smoke tests generate their PDFs and reach their pipeline assertions on supported Windows and Linux environments.

Written by the indexing model from the issue text.