Fails to properly extract table in official Swiss certificate of pension contributions

### Bug
Docling is hardly able to detect and extract tables in the given PDF form.

Input looks like:

And the Markdown result renders one single table with a wrong last row.
[file.md](https://github.com/user-attachments/files/30708461/file.md)

The hybrid chunker then even breaks that single table apart into meaningless values.

### Steps to reproduce
1. Launch docling-serve-cu130:v1.27.0
2. Connect to UI (e.g. http://localhost:5001/ui/)
3. Select https://www.estv.admin.ch/dam/de/sd-web/SGAqZmbwA2OC/21EDP-2026-de.pdf
4. Keep default settings, and run process
(chunking cannot be done via UI, but that's anyway not part of this report).

### Docling version
2.113.0 running in docling-serve-cu130:v1.27.0

### Python version
Python 3.12.13 in docling-serve-cu130:v1.27.0

Reproduce the report with docling-serve-cu130:v1.27.0 using the linked Swiss PDF and compare the Markdown output with file.md. Start from the UI processing path, then follow the hybrid chunker behavior described in the issue. Done means the form's tables are extracted with the correct rows and are not split into meaningless values by the hybrid chunker.

Written by the indexing model from the issue text.