[FEATURE] Add performance mode

[hughperkins]

Description

• Compilation consistently takes around 30s on Linux (Coreweave NVIDIA H100)
  • This translates in x1.4 speedup for MPR, x2 speedup for MPR+SDF, and x6 speedup for GJK
• Runtime speed is within ~1%-5% slower

Related Issue

Resolves #1329

Motivation and Context

As a simulation user,
I want to be able to compile Genesis quickly
Because I don't want to wait for it to compile.

How Has This Been / Can This Be Tested?

Run franka arm with/without the option, after deleting the cache each time. In more detail:

• clone or save https://github.com/hughperkins/taichi-play/blob/13dfd64a1f77e46419de1bf63c8f4b863c54eb5c/ndarray_perf/single_franka_envs.py
• run like:

rm -Rf ~/.cache/taichi/; time python ndarray_perf/single_franka_envs.py --horizon 100

Result, on 5090:

iteration time 668.4088706970215 us
[Genesis] [18:54:24] [INFO] 💤 Exiting Genesis and caching compiled kernels...

real    0m26.231s
user    0m37.009s
sys     0m1.595s

compare with:

rm -Rf ~/.cache/taichi/; time python ndarray_perf/single_franka_envs.py --horizon 100 --performance-mode 1

Result, on 5090:

iteration time 661.4446640014648 us
[Genesis] [18:58:11] [INFO] 💤 Exiting Genesis and caching compiled kernels...

real    2m8.540s
user    2m19.173s
sys     0m1.731s

Screenshots (if appropriate):

Checklist:

• I read the CONTRIBUTING document.
• I followed the Submitting Code Changes section of CONTRIBUTING document.
• I tagged the title correctly (including BUG FIX/FEATURE/MISC/BREAKING)
• I updated the documentation accordingly or no change is needed.
• I tested my changes and added instructions on how to test it for reviewers.

• I have added tests to cover my changes.
• All new and existing tests passed.

[duburcqa]

Make sure that the benchmarks are still using performance mode, and add one benchmark in the parameter matrix that check the performance for non-performance mode.

[hughperkins]

Make sure that the benchmarks are still using performance mode, and add one benchmark in the parameter matrix that check the performance for non-performance mode.

Per out-of-band discussions, we will benchmark only with performance mode off, and then you will look at creating a new bencharmk CI, that runs once a ~week, that will run with a more diverse set of parameter combinations.