dropping direct cuda support

[cwsmith]

CUDA 13 appears to break building Omega_h with the direct CUDA backend:
https://github.com/spack/spack-packages/pull/2059/files

Given that, I think we should drop support for that backend and require Kokkos for CUDA support.

Omega_h can work with Intel, AMD, and NVIDIA GPUs via Kokkos without the use of any vendor specific API calls. The long term plan was to drop support for the vendor specific backends in favor of using Kokkos. This is the first issue that has motivated the change.

This could be combined with the change to the default tag type as a major version increase.

Any thoughts or objections @joshia5 @jacobmerson @Sichao25 @Angelyr ?

[jacobmerson]

One thing to double check is that R3D works without the CUDA backend. See here: SCOREC/pcms#209 (comment)

[cwsmith]

Yeah. There should be no major issue keeping it working. There is a test (

omega_h/src/CMakeLists.txt

Lines 518 to 519 in ca63940

	osh_add_exe(r3d_test)
	test_basefunc(run_test_r3d 1 ./r3d_test)

) and we had it working with Kokkos+SYCL and Kokkos+HIP prior to this work.

[joshia5]

Given that, I think we should drop support for that backend and require Kokkos for CUDA support.

@cwsmith 👍 , iirc it's only a few seconds added overhead during building omegah with kokkos, instead of just loading cuda/vendor-specifc modules

[cwsmith]

closed with #181