Fix mixed image channel sizes in image classifiers

[Creylay]

Summary

Image classifiers (MLP, CNN, LeNet5) failed when training on datasets containing images with mixed channel formats and/or class-imbalanced splits. Two independent bugs were fixed:

1. The transform pipeline did not normalize image channels before converting to tensors, causing torch.stack to fail when a batch mixed grayscale, RGB, and RGBA images.
2. label_to_idx was built only from training split labels, so classes with no training examples (present only in validation) had no index. This caused prepare_output to produce a mixed int/string list that PyArrow could not convert.

---

Type of Change

• Backend change
• Frontend change
• CI / Workflow change
• Build / Packaging change
• Bug fix
• Documentation

---

Changes (by file)

• DashAI/back/models/mlp_image_classifier.py: added convert("RGB") to transform pipeline; build label_to_idx from Categorical.categories (full dataset class list) instead of training split labels; changed prepare_output fallback from string to -1.
• DashAI/back/models/cnn_image_classifier.py: same fixes.
• DashAI/back/models/lenet5_image_classifier.py: same fixes.
• DashAI/back/models/base_torchvision_image_classifier.py: label index fixes only (convert("RGB") was already present).

---

Testing

Train any image classification model on a dataset that:

• Mixes grayscale, RGB, and/or RGBA images — training should complete without RuntimeError: stack expects each tensor to be equal size.
• Has classes with few examples that may land entirely in the validation split — training should complete without ArrowInvalid: tried to convert to int64 or y_true contains values {-1}.