Reduce host memory usage by storing images as uint8

[miakh]

Most input images are 8-bit, but OpenSplat currently keeps them in memory as float32. This change stores loaded images and image pyramids as uint8, converting them to float only when they are used.

In a quick RTX 3090 test, peak RAM dropped from 7.85 GiB to 2.67 GiB (66%) with no meaningful change in runtime or loss.

Tested on macOS/MPS and Linux/CUDA.

Addresses #134.

[pierotofy]

What was the runtime difference? That's a lot of conversions.

[miakh]

This reduces host RAM, not GPU VRAM. Images are stored as uint8 in RAM and converted to the same float32 GPU tensors as before. With GPU caching, this normally happens only once per image. So there are no additional conversions. On the RTX 3090 test - GPU VRAM stayed at ~6.5 GiB.

Runtime was 19.22 vs. 18.81 seconds, so effectively unchanged.

On the M5’s unified memory, usage dropped from 12.08 to 9.49 GB (-21%) and runtime from 52.35 to 46.53 seconds (-11%), so - yea speedup but rather due to lower memory pressure.

Tested on the Person Hall dataset: 314 images at 5616×3744, downscaled to 1404×936 for training.

[pierotofy]

Please don't submit any further pull requests. 🙏