Use the taichi autodiff grad during backward

[zjcs]

When I read the taichi doc, taichi support autodiff funcation by kernel_func.grad. So is it possible to directly use kernel_func.grad in the backward function during taichi-3dgs?

[wanmeihuali]

I've tried that before: #53
It seems taichi's autodiff system is still buggy and cannot produce the correct gradient. It seems the taichi's autodiff implementation is pretty naive.

[oliver-batchelor]

It works OK for simple things, for example, projection or evaluating spherical harmonics - I have used it in my implementation (based originally on this one). https://github.com/uc-vision/taichi-splatting

But it certainly isn't totally straightforward; working around bugs can be painful, and you can't hope to use it in places like the rasterization loop. In my opinion, it would have been better if they had implemented the autodiff function.