Fix knn_graph self-loops with coincident points and shared network nodes

[pablo-legerensomolinos]

Summary

knn_graph dropped each point's own entry from its neighbour list by position
(the first result) instead of by index. When several points share a location, or
snap to the same network node, all distances are 0, so a point is not necessarily
returned first. This produced self-loops (u, u) and dropped a real neighbour.

The point is now removed by index in both the euclidean/manhattan branch and the
network branch. Behaviour is unchanged when no ties are involved.

Related issues

Closes #231

Testing

• uv run pytest -q (Python 3.14.3, MPLBACKEND=Agg): 941 passed, including
  tests/test_graph.py; proximity.py at 100% coverage. The 3 new regression tests
  fail without the fix. Also run on Python 3.12 (805 passed, excluding
  tests/test_graph.py).
• uv run pre-commit run --all-files: all hooks pass. I run Windows with
  core.autocrlf=true, so yamllint only passes once the YAML files have LF endings;
  this is unrelated to the change.
• Without MPLBACKEND=Agg, two plotting tests failed intermittently on my machine with
  a Tk TclError; they pass in isolation and with Agg, and the change does not touch
  plotting code.
• Compared against a brute-force kNN on random data: 600 euclidean/manhattan trials
  (half with forced coincident points) and 120 network trials on a grid network show no
  discrepancies (the previous code showed 266 and 99).
• Not run locally: Python 3.13 (left to CI).

Documentation

Not applicable.

Reviewer notes

• Tests go through the public knn_graph API only.
• The network branch now uses a stable sort so ties resolve deterministically by
  node position.
• Not covered here: delaunay_graph, gabriel_graph, relative_neighborhood_graph
  and euclidean_minimum_spanning_tree raise QhullError for collinear points; I'll
  raise that separately.

[pablo-legerensomolinos]

Hi! The two workflows on this PR are waiting for maintainer approval ("action_required", since this is my first contribution from a fork). Could someone approve and run them when convenient? I ran the full suite locally (941 passed, incl. tests/test_graph.py`) and all pre-commit hooks pass, but Python 3.13 and the docs build are only covered by CI. Thanks!

[yu-ta-sato]

Hi @pablo-legerensomolinos, thanks so much for raising this PR! It totally makes sense and yes the algorithm relied on the position-based identification which should have been improved! Let me see the scripts and run them on my env, and if no problem let me merge it immediately!

[yu-ta-sato]

I don't think there's any issue regarding your approach. Thanks so much, let me merge it to the main branch :)