[pFornagiel]

This PR adds a second 2.5D segmentation network based on architecture suggested by nnUnet plans with mobile blocks and makes it the default. The existing 3-channel MobileNetV3 U-Net stays available, and which model runs is chosen with an environment variable.

Summary

• Renamed the previous architecture to MobilenetV3Unet3Ch
• Added new default model: LiteMobileUNet5Ch, 5-channel U-Net with a MobileNet-style encoder, modelled after nnUnet plans based on the dataset. Seven encoder stages of plain inverted residuals (blocks mobilenetv3 uses) go down to 1/64 resolution, and the decoder uses depthwise-separable blocks. Each slice is stacked with two pairs of neighbours, at ±9 and ±27 slices (SliceContext25D5Channel, edges clamped). Z-score normalisation runs on the volume before that stacking, so all five channels share one set of statistics.
• The rest of the pipeline is unchanged: resample to 0.3 mm, reorient to RPI, paddubg, sliding-window 2D prediction, then reverse pad, drop labels above 32, remove components under 40 mm^3, and resample back to the input grid.

Selecting a model

Model can now be configured through the environmental variables:
TOOTH_MODEL picks lite_mobile_unet_5ch (default) or mobilenetv3_unet_3ch.
TOOTH_MODEL_PATH still points at an ONNX file and overrides the above.