Revert "Fix dtype bug in image converter (#2147)"

[mattdangerw]

This reverts commit e9241dd.

@abheesht17 reverting this because one of the main modes of operation is taking a uint8 image -> rescaled float output. Basically keras_hub.layers.ImageConverter(scale=1./255)(int_image). Whit this change we would cast back to int and flatten the image into all 0s and 1s.

[abheesht17]

This reverts commit e9241dd.

@abheesht17 reverting this because one of the main modes of operation is taking a uint8 image -> rescaled float output. Basically keras_hub.layers.ImageConverter(scale=1./255)(int_image). Whit this change we would cast back to int and flatten the image into all 0s and 1s.

This will break PaliGemma though. At least, I was not able to run it on Colab (maybe, doing something wrong). Here's the notebook: https://colab.research.google.com/gist/abheesht17/c4eb5ecf392143ba8927c116954b8aa8/kerashub-image-converter-dtype-bug.ipynb. So, we should fix that?

[mattdangerw]

Thanks! I'll take a look. Right now the basic image backbones (resnet, etc.) are broken, but let's find a way that works for everything.

[abheesht17]

Oh, and Gemma3 will break too

[mattdangerw]

@abheesht17 I think this fixes things. Can you check?

[mattdangerw]

Ugh this is whack-a-mole. Take two.

[abheesht17]

Thanks! Working with Gemma3. Merging