Unable to convert "large" PDFs, silent failures or ERROR docling.pipeline.standard_pdf_pipeline: Stage preprocess failed for run 1, pages [80]: std::bad_alloc

### Bug

I have an 80Mb PDF, which has 184 pages of scanned grayscale images.

Using:

docling --device cuda --to html_split_page --image-export-mode embedded Manual.pdf

Will result in a steady number of `std::bad_alloc` errors.

1. Using `--pdf-backend pypdfium2` avoids this problem completely.
2. Using `--pdf-backend dlparse_v4` similar error as default docling_parse
3. Using `--pdf-backend threaded_docling_parse` silently fails with no output to disk

I get the impression from:

* https://github.com/docling-project/docling/issues/3345
* https://github.com/docling-project/docling-parse/issues/227

That this maybe something under investigation but both issues are closed.

I think the pypdfium2 workaround works OK but comments in https://github.com/docling-project/docling-parse/issues/227#issuecomment-4652392034 suggest there maybe differences in quality depending on the backend

Questions:

1. does changing the default make sense?
2. would supporting batch mode for pages work? I tried `--release-native-memory-every-n-pages 20` with no success
3. would start/end pages support help here as an alternative workaround?

### Steps to reproduce

See bug description

### Docling version

Docling version: 2.104.0
Docling Core version: 2.83.1
Docling IBM Models version: 3.13.3
Docling Parse version: 6.2.0
Python: cpython-312 (3.12.1)
Platform: Windows-11-10.0.22631-SP0

### Python version

Python 3.12.1

This seems redundant, see Docling version above

Start by reproducing the 184-page conversion with the documented docling command and compare the default, pypdfium2, dlparse_v4, and threaded_docling_parse backends. Investigate the reported std::bad_alloc on page 80 and the silent no-output behavior, including whether page batching or start/end page options changes the result. Done means large PDFs complete reliably or fail with a visible, actionable error.

Written by the indexing model from the issue text.