Upgraded resize_image function to support interpolation

[JTStephens18]

Replaced convolution operation with interpolation. Added support for downscaling using floats.

resolves #3263

[jb-ye]

I am not in favor of using aliased downscaling method for training GS. See my response here: #3263 (comment)

[kerrj]

One option is to use torchvision's resize which has antialiasing?

[brentyi]

Similarly, can we just set antialias=True in torch.nn.functional.interpolate()?

[jb-ye]

One option is to use torchvision's resize which has antialiasing?

I tried this before, it still doesn't bring the same PSNR results as convolution (that's why I eventually changed it to convolution). I am a little unsure if those antialiased implementation in pytorch is correct at all. For example, as noted by this paper, most existing downscaling implementation doesn't implement antialiasing properly.

[jb-ye]

Similarly, can we just set antialias=True in torch.nn.functional.interpolate()?

I am curious to see results on some MipNerf360 datasets and see whether it degrades the performance by checking the the PNSR eval curve. The current choice was made based on my observation of bicycle and garden dataset.

See #2984