feat: nnUNet integration for CoPick cryo-ET datasets

[jtschwar]

Adds a full nnUNet training and inference pipeline accessible via the copick-torch CLI. The workflow covers three stages — dataset preparation, model training, and inference — each exposed as a
CLI subcommand and registered as a copick entry point.

What's new

copick_torch/nnunet/ module

• prepare (copick.convert.commands): Converts a CoPick project into nnUNet raw dataset format. Reads tomograms and multilabel segmentations, writes them as .nii.gz NIfTI files in the imagesTr /
  labelsTr / imagesTs directory structure, and emits a dataset.json. Voxel spacing is converted from Ångströms to nanometres for nnUNet's patch-size planner. Conversion is multi-threaded.
• train (copick.training.commands): Runs nnUNetv2_plan_and_preprocess followed by nnUNetv2_train. Supports the standard nnUNet trainer, the Residual Encoder Large variant (resnecl), and all ten
  MedNeXt variants (S/B/M/L × kernel-3/kernel-5). Multi-GPU DDP is supported via --num-gpus; the plans batch size is scaled automatically so per-GPU memory stays constant. Resumes from an existing
  checkpoint if one is found.
• segment (copick.inference.commands): Wraps nnUNet inference as nnUNetPredictor. Single-volume inference runs on cuda:0; batch inference shards runs round-robin across all available GPUs using
  mp.spawn. Supports fold ensembling (pass multiple -w checkpoint paths). Writes predictions back to the CoPick overlay as multilabel segmentations and saves a YAML parameter log.
• mednext_trainer.py: MedNeXt trainer class definitions (10 variants). Copied into nnUNet's trainer discovery directory on first use so the standard nnUNetv2_train CLI and checkpoint loading
  both work without patching nnUNet.

pyproject.toml

• Adds nnunetv2 and SimpleITK as dependencies.
• Registers CLI entry points in copick.convert, copick.training, and copick.inference command groups.

Usage sketch

1. Prepare dataset

  copick nnunet prepare -c config.json --tomo-uri wbp@10.0 \                                                                                                                                  
      --seg-info targets --dataset-name CZII --raw /data/nnunet_raw

2. Train (standard nnUNet, fold 0)

  copick nnunet train -n CZII -r /data/nnunet_raw \                                                                                                                                           
      --preprocessed /data/preprocessed --results /data/results     

3. Segment

  copick nnunet segment -c config.json \                                                                                                                                                      
      -p /data/results/.../nnUNetPlans.json \                                                                                                                                                       
      -d /data/nnunet_raw/Dataset001_CZII/dataset.json \                                                                                                                                            
      -w /data/results/.../fold_0/checkpoint_best.pth