[CSSFrancis]

Adds Direct Electron's diffraction spot (Bragg disk) networks as twelve kind: weights entries (each network as an ONNX graph and as a torch checkpoint), served from Zenodo record 23383572 (DOI 10.5281/zenodo.23383571), licensed CC BY-NC 4.0.

Entry	File	Role
DE_SpotNet	de_spotnet.onnx	default spot detector (scale-robust)
DE_SpotNet_Dynamical	de_spotnet_dynamical.onnx	detector fine-tuned on dynamical disks, with a disk-coverage head
DE_SpotNet_Production	de_spotnet_production.onnx	detector variant, higher recall at the canonical disk size
DE_SpotNet_Base16	de_spotnet_base16.onnx	earlier detector, kept as a fallback
DE_CenterNet	de_centernet.onnx	re-places each detected disk on the raw pattern; gives a positional sigma
DE_FriedelNet	de_friedelnet.onnx	Friedel-partner refinement after DE_CenterNet

• The six entries above point at ONNX graphs (model.framework: onnx, model.class: onnxruntime.InferenceSession), so pyxem and quantem can run them without torch.
• Each has a _Torch partner (DE_SpotNet_Torch, ..., DE_FriedelNet_Torch) for the .pt checkpoint in the same record (model.framework: torch), which training and fine-tuning start from. The descriptions cross-reference each pair.
• Checksums and sizes are filled in (this comes from a fork, so the fill-in workflow cannot push).
• One dated version each, "261015".
• Used by pyxem (Add get_diffraction_vectors(method="neural") on ONNX Runtime with an anywidget GUI, and device= for faster template matching pyxem/pyxem#1194, get_diffraction_vectors(method="neural")) and quantem (BraggVectors.detect_disks(method="neural")).

Checked locally: --validate passes for all twelve, the stub is regenerated (_create_stubs --check clean), the test suite passes (376 passed), ruff and basedpyright are clean, and every entry downloads through emdatabase.data.<Entry>().download(version="261015") with its md5 verified; the ONNX graphs run in ONNX Runtime.

The second commit makes dataset_files sort on the POSIX path string: Windows Path objects compare case-insensitively, so a stub generated on Windows put LayeredCuNb4DSTEM before LSMOLineScan and failed the stub check on Linux.

[CSSFrancis]

@arthurmccray do you want to review this? Not sure exactly how we should be distributing these. The ONNX models are nice so they aren't so tied to pytorch

[arthurmccray]

yup I'll try and get to it today!

[arthurmccray]

lgtm, zenodo is down (again) but I trust all the links will work when it's back up