Drop vendored nnunetv2; depend on upstream pip package (v1.3.0)

[77even]

Summary

BREAKING CHANGE — drops the vendored distillation/nnunetv2/ tree (~190 files, 1.7 MB) and switches to a plain pip dependency on nnunetv2>=2.5,<3.0. Bumps the package version to v1.3.0.

• The vendored snapshot had drifted from upstream and there was no automated way to refresh it.
• We had only ever authored one file inside that tree (nnUNetDistillationTrainer.py); the other ~190 files were stale upstream code.
• That single file is now distillation/nnunet_distillation_trainer.py (top-level, next to the other scripts).
• The sys.path.insert hack that forced the vendored copy to win over pip has been removed from all entry-point scripts.

Rebased onto main (post-v1.2.4)

This branch has been rebased onto main after v1.2.3 (PR #6) and v1.2.4 (PR #7) landed. v1.3.0 therefore includes everything from both:

• From v1.2.3 — 2d, 3d_lowres, and 3d_cascade_fullres configurations work end-to-end (was hardcoded 3D-only before). ONNX export adapts dummy input shape and dynamic axes per configuration dim. Output filenames include the configuration name so different configs don't overwrite each other.
• From v1.2.4 — Primus knowledge-distillation support for upstream nnUNet_Primus_{S,B,M,L}_Trainer teachers. New nnUNetv2_primus_distillation_train and nnUNetv2_primus_distillation_export_onnx entry points; new primus_distillation_trainer module.

After rebase, primus_distillation_trainer.py's import of the distillation trainer was updated to point at the unvendored top-level module (see the third commit on this branch).

Compatibility fix surfaced by the refactor

nnunetv2 2.7 renamed nnUNetLogger -> LocalLogger. The trainer's import and instantiation have been updated. (This drift was previously hidden by the vendored copy.)

ONNX export prerequisites

PyTorch >= 2.6 changed its ONNX exporter to require onnxscript. Addressed by:

• setup.py: added extras_require={'onnx': ['onnx', 'onnxruntime', 'onnxscript', 'onnx-simplifier']}
• docs/Distillation.md: prerequisites callout before the ONNX export section

Install paths after this PR:

# Training only
pip install -e .

# Training + ONNX export (recommended)
pip install -e '.[onnx]'    # quotes needed for zsh

Commits on this branch

1. Drop vendored nnunetv2; depend on upstream pip package (v1.3.0)
2. Document ONNX prerequisites and add [onnx] extras_require
3. Update Primus trainer imports for the unvendored layout

Test plan

• python -m py_compile passes on all touched files
• After pip install -e '.[onnx]' in distillation/, all six console entry points run --help against pip-installed nnunetv2 2.7.0 (the three existing ones plus the two new Primus ones)
• nnunet_distillation_trainer exposes nnUNetDistillationTrainer, nnUNetDistillationTrainerDA5, LiteNNUNetStudent, LiteResEncStudent
• primus_distillation_trainer exposes nnUNetDistillationPrimusTrainer, nnUNetDistillationPrimusTrainerDA5, LitePrimusStudent, reduce_primus_dims
• CPU smoke test (local-only, not committed) — 11/11 pass: 3D/2D forward+backward for both U-Net student types, cascade-shaped input, distillation loss, 2D ONNX round-trip, Primus reduce-dim invariants, Primus forward/backward, Primus patch_size validation, Primus ONNX round-trip
• TODO before merging: GPU end-to-end on a real dataset for each configuration / teacher type:
  • U-Net student: 3d_fullres, 2d, 3d_cascade_fullres. Cascade requires upstream nnUNetv2_train ... 3d_lowres + nnUNetv2_predict --save_probabilities so predicted_next_stage/ exists.
  • Primus student: at least one of -ts S / M / L with -r 2. Primus teacher must be trained first with stock nnunetv2 (nnUNetv2_train DATASET_ID 3d_fullres 0 -tr nnUNet_Primus_M_Trainer).
• TODO before merging: ONNX export end-to-end against a real trained checkpoint (U-Net and Primus), including the --simplify path.

Upgrade notes for users

git pull
pip install -U -e distillation/                # core only
pip install -U -e 'distillation/[onnx]'        # with ONNX export

Existing distilled checkpoints continue to load; the trainer/student APIs are unchanged from v1.2.4.