Implement a detail variation method of the image input

[FishWoWater]

According to Sec 3.4 of the paper, I have implemented the detail variation method given the image input.

Pipeline:

1. Given a Z-up axis triangle mesh, it is voxelized(https://github.com/Forceflow/cuda_voxelizer) into active voxels (64^3 -> N x 4), in the same form of Sparse Structure Voxels
2. Skip the sparse structure sampler and directly run the 2nd stage, SLAT sampler, based on the input image condition.

I have provided a example script, texgen tab, as well as example data

It can composite meshes and images of different styles, producing some creative outputs.

Example outputs:

typical_humanoid_block_robot+typical_humanoid_mech_gs.mp4

typical_creature_dragon+typical_creature_furry_gs.mp4

typical_creature_dragon+typical_creature_elephant_gs.mp4

Also other minor changes:
i.e. options to save intrinsics/extrinsics/renderings in the example
output sRGB texture
expose opt/fast mode in texture baking

[FishWoWater]

@microsoft-github-policy-service agree

[YuDeng]

Nice implementation! The image-conditioned variation results are really cool! We have only tested text-conditioned variation in our experiments. This one seems to be more creative.