Bugs when preprocessing the depth of Hypersim

[thucz]

https://github.com/nihalsid/panoptic-lifting/blob/b8ad4f5941ed0158f365d2c19facc86fbae2061e/dataset/preprocessing/preprocess_hypersim.py#L420

I found the maximum depths of some scenes in Hypersim after multiplying 1000 are far larger than the maximum value of int16(32767). It is not adequate to use int16 to compress the depth of Hypersim. I think just using .npy format with float32 is OK.