fix: address issue #1948

[ghost]

Fixes #1948

What was broken

When NaturalImage2DIO was used for 2D PNG datasets, label files saved as
RGB (3-channel) PNGs were silently loaded as (3, 1, H, W) instead of
(1, 1, H, W). This passed verify_dataset_integrity (which only
compared the spatial dimensions of images and segmentations and never
checked that the segmentation itself was single-channel) and only
exploded later during training inside the data loader with the cryptic
error:

could not broadcast input array from shape (3,301,301) into shape (1,301,301)

raised at nnunetv2/training/dataloading/data_loader_2d.py:92 when the
seg array was padded into the pre-allocated single-channel seg_all
buffer.

What this fixes

Two minimal, targeted changes make the failure happen early with a clear
message instead of crashing mid-training:

1. nnunetv2/imageio/natural_image_reader_writer.py — NaturalImage2DIO.read_seg
   now checks the channel dimension of the loaded segmentation and
   raises RuntimeError with a message that explains the cause and the
   fix (re-save labels as single-channel grayscale images with integer
   class indices) when a multi-channel label file is encountered.

2. nnunetv2/experiment_planning/verify_dataset_integrity.py — check_cases
   additionally asserts segmentation.shape[0] == 1 and prints a clear
   error pointing at the offending label file, so verify_dataset_integrity
   now actually catches this dataset issue instead of giving a false
   "Done" and letting training crash later.

No behavioral change for users with correctly formatted single-channel
labels.

How to verify

1. Build a tiny 2D PNG dataset of the kind that triggers the bug
   (any 2D dataset using NaturalImage2DIO) and save one of the label
   PNGs as RGB (e.g. with Pillow's mode="RGB").

2. Run nnUNetv2_plan_and_preprocess -d <ID> --verify_dataset_integrity.
   With the fix, verify_dataset_integrity fails immediately with a
   message like:

   Error: Segmentation must be a single-channel integer label map but has 3 channels. ...
   Seg file: <path>
   

   instead of printing verify_dataset_integrity Done. and proceeding
   into preprocessing/training before crashing inside the data loader.

3. Re-save the label as a single-channel mode="L" PNG and confirm the
   pipeline runs through preprocessing as before — no regression for
   well-formed datasets.