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.