ND, feature transform, tests, performance improvements

[kevinjohncutler]

Summary

• Sync legacy sources to upstream commit 672a57c (legacy matches upstream exactly).
• Add ND feature transform tests (brute-force correctness + distance consistency + dtype/edge cases).
• Remove temporary upstream snapshot folder.

Testing

• PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 python -m pytest -v

Notes

• Legacy vs ND ratios remain < 1 in recent bench runs (see benchmarks/legacy_vs_nd_explicit_*.csv if needed).

[kevinjohncutler]

Ok, I have completed my cleanup of the code and addressed all the above concerns and ideas. Ready for review! If everything else looks good, we can strip out the legacy comparisons as a final step.

[william-silversmith]

Thank you! I'll dig into the core logic some more. Thank you for all the work you've put into this PR so far!

[kevinjohncutler]

@william-silversmith I think I figured out a way to generalize to any L-norm. Interested in making that part of this PR? Might have to rebrand to "efficient distance transform" or "exact distance transform" haha

[william-silversmith]

That's a very interesting idea (it's discussed in the Felzenszwald and Huttenlocher paper I believe), but let's limit the scope of this PR, it's very complex already containing at least 4 different ideas and a full rewrite. Once it goes through, it can be built off of. Appreciate your enthusiasm!

[kevinjohncutler]

I happened to drum up an implementation of expand_labels with numba that beat the implementation we had in edt. However, I was able to learn from and surpass the numba implementation, which happened to give the main edt calls a speed boost as well (1.2x on average).

[william-silversmith]

Sorry for the delay! Will get to reviewing the core very soon. Just been busy.