Fix shape mismatch

[Laughing-q]

@AyushExel @glenn-jocher
the shape mismatch issue happens when there're mixed labels(both detection labels and segmentation labels).
Now it will force to use detection labels and remove segmentation labels when there're mixed labels and raise a warning.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhancements in label caching and consistency checks for mixed datasets.

📊 Key Changes

• 📝 Improved the cache_labels method to update im_files based on the cached labels.
• 🚀 Added consistency checks to handle datasets containing a mix of bounding boxes and segments within the get_labels method.
• ⚠️ Added a warning for when a dataset contains both detection boxes and segmentation labels, now defaults to using only boxes in such cases.

🎯 Purpose & Impact

• 🎨 The update ensures the list of image files remains consistent with the cached labels, improving data integrity.
• 💡 This change helps prevent potential issues when training models on mixed datasets, by enforcing dataset consistency and creating awareness about the ideal dataset structure.
• 🛠️ Benefits users by automatically resolving inconsistencies in their data which might have led to training errors or suboptimal model performance.

[Laughing-q]

tested this dataset with the same command and it works as expected. #548 (comment)
before:

after this PR, the training looks good and raise a warning says the dataset is not pure:

[glenn-jocher]

@Laughing-q PR merged! :)