Fix invalid DOI and strengthen bibliography references

[theo-barfoot]

Context

Raised during JOSS review (openjournals/joss-reviews#10072) by editorialbot and @yewentao256.

Problems

1. Invalid DOI flagged by editorialbot

10.5555/3327757.3327857 is INVALID — 10.5555 is a known broken prefix, replace with https://dl.acm.org/doi/{doi} in the url field.

This DOI corresponds to the PyTorch NeurIPS 2019 paper. The 10.5555 prefix is a pseudo-DOI from ACM that doesn't resolve. If this DOI is still present in paper.bib, remove it and use a URL to the official proceedings page instead.

2. Missing DOIs (skipped by editorialbot)

The following references had no DOI and editorialbot couldn't find one automatically:

• PyTorch
• CuPy
• JAX
• Fast Graph Representation Learning with PyTorch Geometric
• Solving sparse linear systems in PyTorch

For each, verify whether a DOI exists and add it if so. For software/blog references where no DOI exists, ensure a url field is present.

3. Add formal citations for PyTorch issues mentioned in text

@yewentao256 recommended strengthening references by "adding formal citations/URLs for the specific PyTorch issue(s) mentioned and for upstream code sources referenced in the text/acknowledgements."

The paper currently references PyTorch issue #41128 inline. Rather than adding GitHub issues as bib entries, consider:

• Keeping inline references as "PyTorch issue #41128" with a footnote URL
• Adding citations for PyTorch sparse/masked tensor documentation where relevant
• Ensuring upstream code sources acknowledged in the Acknowledgements section (pykrylov, cornellius-gp/linear_operator, pytorch-minimize) have proper citations or URLs

Acceptance Criteria

• Invalid 10.5555 DOI removed or replaced with a working URL
• All references have either a valid DOI or a url field
• PyTorch issue references are clearly linked (footnote or inline URL)
• Upstream code sources in Acknowledgements have proper attribution