Create NVRTCC, a utility for running tests under NVRTC

[wmaxey]

Description

Closes #465

This PR allows us to test NVRTC just like any other compiler. It provides compile time feedback and enables negative compilation tests. e.g. compile.pass, .fail, or possibly even weird XFAIL cases.

TODO:

• Implement NVRTCC
• Connect to lit
• Document / update CI to build the libcudacxx project before executing lit.

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[wmaxey]

Current status on Windows:

PS C:\> lit -sv -j20 .\cccl\libcudacxx\.upstream-tests\test -Dcompute_archs=80
...
Failed Tests (2):
  libcu++ :: std/containers/views/mdspan/mdspan.extents.cons/array.pass.cpp
  libcu++ :: std/containers/views/mdspan/mdspan.extents.cons/span.pass.cpp


Testing Time: 276.21s
  Unsupported      :  459
  Passed           : 1257
  Expectedly Failed:    5
  Failed           :    2

[wmaxey]

We can now build NVRTC tests with -Dexecutor='NoopExecutor()' to disable execution if we wanted to, though I believe it still requires a GPU or the CUDA driver installed. Not sure which.

[jrhemstad]

snarky suggestion: std::regex. I'm not even mad, I'm impressed. However, would it be easier to just use an existing lib like cxxopts for this?

[wmaxey]

I think as we discussed. Pulling in dependencies where we might need to be building in 'weird places' makes me not want to use them... STL-only seems safest from a 'will work almost anywhere' standpoint.

[jrhemstad]

question: Instead of needing to pass the produced fatbin to nvrtcc in order to execute the program, couldn't nvrtcc just produce a binary that does this step for you?

Basically, it would store the fatbin in the binary, and in the main() it would do load_and_run_gpu_code(/*stored fatbin*/).

That way, nvrtcc would work even more like a regular compiler in that it outputs a normal binary that can be executed just like ./a.out.

[wmaxey]

This is essentially what the old script did. It produced a binary with the raw test stored inside and compiled it during execution. Every test was recompiled into the same hosting application. It bought us the 'works like a binary thing' but failed for giving us negative testing with .fail.cpp tests.

It might be possible to do this. It would get rid of needing to use the PrefixExecutor() for tests. I'll make a PR after this one with the documentation in mind.

[miscco]

I mean its not pretty, but in that context nothing is.

Great work on improving our test coverage 🎉