[Sparse] Make ti.append() support dynamic snode with **multiple child**

[archibate]

Concisely describe the proposed feature
Currently ti.append(snode, n, val) only supports dynamic snode with exactly one child.
I think why @yuanming-hu do such design is limited by the fact that it's hard to make val extendable.

Describe the solution you'd like (if any)
Instead of passing val to append, how about this: we just allocate a slot for it, but not initialize it with val. The initialization can be done later since append is atomic.

---

Old API:

ti.dynamic(ti.i, 233).place(ch1)

ti.append(ch1.snode(), [], val)  # old ad-hoc inline value appender

New API:

ti.dynamic(ti.i, 233).place(ch1, ch2, ch3)

n = ti.append(ch1.snode(), [])  # atomically increase dynamic snode length by 1
ch1[n] = n  # since n is atomic, so it's fine to let the user fill the slots one-by-one
ch2[n] = n * 2
ch3[n] = n * 3

Additional comments
I found this issue when developing https://github.com/taichi-dev/taichi_three with a OpenGL-alike API glQuad and glSphere, which needs to dynamically append triangle vertices.


(Sorry, I can't help but show off the cool works in Tai3D :)
We may want to start solving this issue after #1256 merged to prevent conflict.

[yuanming-hu]

The visual results look good! Btw, the dynamic node is going to be redesigned to support various extensions like the one you mentioned here. I'll share the refactoring plan in the week of June 29.