Add prgenv-gnu-dpcpp: Intel DPC++ SYCL for GH200 aarch64

[dganellari]

Add experimental recipe that extends prgenv-gnu/25.11 with Intel DPC++ (SYCL compiler with CUDA backend) for GH200 on Alps aarch64 systems.

The Intel oneAPI binary distribution is x86_64 only, so DPC++ is built from source using a custom Spack package (llvmdpcpp) that fetches the intel/llvm sycl branch.

Changes:

• New recipe at recipes/prgenv-gnu-dpcpp/25.11/gh200/
• Custom Spack package repo/packages/llvmdpcpp/package.py
• Spack v1.1.1, spack-packages 7178c89 (2026-03-09)
• Registered in config.yaml (build only, no deploy)

Tested on Santis and Clariden up to 2262 nodes (9048 GPUs).

[bcumming]

cscs-ci run alps;system=daint;uarch=gh200;uenv=prgenv-gnu-dpcpp:25.11

[msimberg]

cscs-ci run alps;system=daint;uarch=gh200;uenv=prgenv-gnu-dpcpp:25.11

[dganellari]

@msimberg seems like it failed at some NCCL test.

[msimberg]

cscs-ci run alps;system=daint;uarch=gh200;uenv=prgenv-gnu-dpcpp:25.11

[msimberg]

Oh, the llvmdpcpp package doesn't pin the commit but just uses the latest commit on the sycl branch. This makes the build very unreproducible, and I'm guessing this is why the build is failing now, but didn't before.

@dganellari do you have a way to find out which commit was actually used for the previous (successful) build? If yes, could you check that and pin the commit? If no, let's go with the last successful build and just deploy that instead (with a big TODO on the version saying it should be pinned).

[msimberg]

I deployed the latest build to daint:

$ uenv image find prgenv-gnu-dpcpp
uenv                       arch   system  id                size(MB)  date
prgenv-gnu-dpcpp/25.11:v1  gh200  daint   e8bc614187bd742a   5,769    2026-06-19

@dganellari will look into the llvmdpcpp version/commit still so let's keep the pr open.