Squash clear_list calls into one kernel?

[k-ye]

Concisely describe the proposed feature

Right now, before launching a kernel that iterates over a sparse tensor, Taichi will generate a series of

clear_list(s1);
element_listgen(s1);
clear_list(s2);
element_listgen(s2);
...

@yuanming-hu I think we can have a single kernel for all the clear_list() calls along the tree?

This idea came from #896 , where the user has noticed a slower performance when using sparse tensors, due to the overhead. I don't think clear_list() is the major overhead here, but having as fewer kernels as possible might always be a good thing..

[yuanming-hu]

Let me tell you a secret: after #742 we will have kernel fusion and this can be done automatically :-)

[k-ye]

Haha great! 😆