[lang] Cleanup parallel sort utility

[AmesingFlank]

This PR moves the parallel_sort function to a different place, so that we don't use import taichi as ti. It also moves the kernel definition outside the python function, so that the kernel isn't recompiled everytime the function is called.

On a tangential note, this sort function is kinda slow. On my macbook, sorting the particles in mpm99 is so slow that the simulation goes down to around 1.2 FPS. This is quite bad, I might try to implement a fast radix sort.

[netlify]

✔️ Deploy Preview for jovial-fermat-aa59dc canceled.

🔨 Explore the source changes: 20dd83a

🔍 Inspect the deploy log: https://app.netlify.com/sites/jovial-fermat-aa59dc/deploys/61c3eb60c860390008cb1065

[qiao-bo]

LGTM!

[ailzhang]

btw I feel it's okay to do import taichi as ti in tools :P Is it the only motivation for this change? ;)
import taichi as ti in taichi/lang might be problematic :P

[AmesingFlank]

btw I feel it's okay to do import taichi as ti in tools :P Is it the only motivation for this change? ;) import taichi as ti in taichi/lang might be problematic :P

In my original code in tools, the @ti.kernel was inside a python function, so the kernel decorator is not used until the function is called. As soon as I move the kernel definition to global scope, I started getting error messages that kernel doesn't exists. It is necessary to move the kernel definition out, because otherwise the kernel recompiles every time parallel_sort is called. So basically, putting a taichi kernel inside tools won't work.