Remove hardcoded class count assertions to support multiple datasets

[Copilot]

The validation dataset assertion hardcoded 10 classes, causing failures when training on stanford-dogs (120 classes). The training dataset had a similar hardcoded assertion with a comment indicating support for both 10 and 120 classes, but enforcement didn't match intent.

Changes

• Remove hardcoded class count assertions for both training and validation datasets
• Add dynamic validation ensuring training and validation datasets have matching class counts
• Preserves support for imagenette/woof (10 classes) and stanford-dogs (120 classes)

# Before
assert len(val_dataset.classes) == 10, f"Expected 10 classes, got {len(val_dataset.classes)}"

# After
num_classes = len(train_dataset.classes)
# ... later ...
assert len(val_dataset.classes) == num_classes, (
    f"Training dataset has {num_classes} classes but validation dataset has "
    f"{len(val_dataset.classes)} classes"
)

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.