Support for 16 bit image training and inference

[amrithc]

I am trying to work on 16 bit geospatial imagery and would like to train the model on 16 bit imagery and run it on it. Even if there is no support for it now, any suggestions on how can I add a hotfix right now to make it work till there is an official feature for the support.

[Cyniikal]

You'll need to edit the base mmcv/mmengine image loading stuff to not assume things are uint8. You should be good to go after that.

A lot of stuff assumes uint8, like a lot of the configs for normalization and many data augmentations, so make sure to edit those as well as needed.

(I've done this before for a private fork for work, it's not hard but there are several little things)