Add docs page on GPU kernel compilation and raising

[gdalle]

Following the docs PR invitation in https://discourse.julialang.org/t/kernelabstractions-cuda-reactant-how-to-get-minimal-working-example/134164/

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 64.75%. Comparing base (b39a1fc) to head (da85ba6).
⚠️ Report is 292 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1920      +/-   ##
==========================================
- Coverage   68.16%   64.75%   -3.41%     
==========================================
  Files         109      124      +15     
  Lines       11779    13439    +1660     
==========================================
+ Hits         8029     8703     +674     
- Misses       3750     4736     +986     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[yolhan83]

Hey all, indeed a doc about raising may be interesting but a kernel part may be as important, I think RArray auto trigger CUDABackend on KA which mean without raising it may only work on CUDA, a thing to test is when not raising, setting the backend of KA by hand and the Reactant backend too but not sure Reactant will be able to send the Metal or AMD device code to the XLA backend anyway.

[gdalle]

So, as I understand it:

• Reactant compilation should work without raise=true, but on my Mac it doesn't
• Reactant compilation with raise=true should work without using CUDA, but on my Mac it doesn't

In light of this, should we wait for this to be fixed or should we document the behavior that we know for sure works across platforms?

[Pangoraw]

Reactant compilation should work without raise=true, but on my Mac it doesn't

#1923 would fix this for KA kernels, it should already work for CUDA.jl kernels.

[gdalle]

Thanks! From what I see, the fix lives in Reactant's CUDA extension, so it will remain necessary to run using CUDA even on platforms where it is not functional?

[Pangoraw]

Yes it uses CUDA.jl to extract the kernel IR. So it will need to be loaded (either by the user or the KA Reactant extension)

we should force load this for you

@wsmoses do you know how we can achieve that?

[gdalle]

By making CUDA a hard dep of Reactant? Why isn't that the case if you need it for kernels?

[gdalle]

I understand the need to keep it as a weak dep though, and I don't think Pkg manipulations should be performed under the hood. So the current error message might be good enough, I was just a bit surprised that installing CUDA on my Mac would actually solve the issue (which it did).

[Pangoraw]

By making CUDA a hard dep of Reactant? Why isn't that the case if you need it for kernels?

We need for CUDA kernels and KA kernels, so it can be fair to require using CUDA even in the case of a non-CUDA setup.

[wsmoses]

@gdalle with the other raising PR in, can you rebase this on that