Remove code paths that depend on RMM_STATIC_CUDART

[robertmaynard]

Description

We can remove the optimizations around CUDA_STATIC_RUNTIME and instead see if the function is already in the process space so that RMM doesn't need to have any build context to run properly

Fixes #1679

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[harrism]

Are existing tests sufficient to cover this?

[robertmaynard]

Are existing tests sufficient to cover this?

I will add explicit tests that control the type of cudart we are using to verify these chages

[robertmaynard]

Did more testing today and realize that more work is needed. With this PR we are getting all the symbols required at link time, but the dlsym with RTLD_DEFAULT is failing for a couple of the use cases. The tests still pass in CI since a libcudart.so exists on the machine.

Need more time to figure out what is going wrong.

[vyasr]

In what contexts is this failing? I assume it's a statically linked case, so the symbols cannot have been (dynamically) loaded with RTLD_LOCAL in a way that would obscure them from this scope.

[robertmaynard]

In what contexts is this failing? I assume it's a statically linked case, so the symbols cannot have been (dynamically) loaded with RTLD_LOCAL in a way that would obscure them from this scope.

That is correct it is the static linking use case the is failing. The tests have the textual entries for the symbol, but the dlsym returns null.

[bdice]

This introduces a bit of complexity that we may be able to avoid. RMM declares its minimum supported CUDA version is 11.4 (this has been true since November 2023). We have required a minimum of at least 11.2, usually 11.4, everywhere I can think of for a long time. I think we can remove the shims for CUDA < 11.2 in this PR.