Add multi-configuration support: 2d / 3d_lowres / 3d_cascade_fullres (v1.2.3)

[77even]

Summary

Previously distillation only worked for 3d_fullres. The student model and ONNX export paths hardcoded Conv3d / InstanceNorm3d and a 5D dummy input shape, so 2d, 3d_lowres, and 3d_cascade_fullres all crashed (issue raised in chat).

This PR makes all four standard nnUNetv2 configurations work end-to-end through the distillation + ONNX export path.

Changes

• Trainer (nnUNetDistillationTrainer)
  • Picks conv_op / norm_op from len(patch_size): 2D configs use Conv2d / InstanceNorm2d, 3D configs use Conv3d / InstanceNorm3d.
  • Cascade (3d_cascade_fullres) input channels come from upstream determine_num_input_channels, which already adds num_classes from the previous stage — no extra plumbing needed.
• Student models (LiteNNUNetStudent, LiteResEncStudent)
  • Default kernel_sizes / strides shape and norm_op class are picked from conv_op.
  • norm_op default is now None with auto-selection. Existing callers that pass it explicitly stay valid.
• ONNX export (fast_nnunet_distillation_export_onnx, fast_nnunet_resenc_distillation_export_onnx)
  • Detects dim from patch_size, builds the right N+C+spatial dummy input.
  • dynamic_axes are emitted per-dim (3 spatial axes in 3D, 2 in 2D).
  • Output filename now includes the configuration (e.g. model_2d_fold0_2x.onnx) so 2d / 3d_lowres / cascade exports don't overwrite each other.
• CLI help text on --configuration now lists all four supported values for the four entry-point scripts.
• Docs (docs/Distillation.md): configurations table, example commands per configuration, cascade workflow note.
• Version bump: 1.2.2 -> 1.2.3.

Test plan

CPU smoke tests (local-only, not committed) — 7/7 pass on macOS:

• 3D LiteNNUNetStudent forward / backward
• 2D LiteNNUNetStudent forward / backward (verifies InstanceNorm2d was chosen)
• 3D LiteResEncStudent forward / backward
• 2D LiteResEncStudent forward / backward
• 3D cascade-shaped input (extra input channels) forward / backward
• distillation_loss_fn KL loss + backward
• 2D ONNX export round-trip (max_diff = 2.6e-06)

GPU validation still needed before tagging the release:

• A few distillation training steps on a real dataset for -c 2d, -c 3d_lowres, and -c 3d_cascade_fullres (cascade requires upstream nnUNetv2_train ... 3d_lowres + nnUNetv2_predict --save_probabilities to be run first so predicted_next_stage/ exists).
• ONNX export end-to-end against a real 2D checkpoint, verifying the file name and tensor shape match expectations.

Follow-ups

• After merge: tag v1.2.3, publish release notes (plain Markdown, no emojis).
• Rebase refactor/drop-vendored-nnunetv2 (PR Drop vendored nnunetv2; depend on upstream pip package (v1.3.0) #4) on top of new main. The same changes apply to the moved-out trainer file; setup.py conflict should be resolved by keeping 1.3.0.