Upgrade modern backend to LLVM 21.1.8 and CUDA 13.3

[brandonros]

Replace the modern LLVM 19 backend feature with LLVM 21 while retaining LLVM 7 for the legacy path. Update feature forwarding, LLVM tool discovery, the vector-add example, and documentation together. The modern Nix shell uses LLVM 21.1.8 and CUDA 13.3; CUDA 13 Docker definitions use CUDA 13.3.1 bases. Update the Windows CUDA installer action to support CUDA 13.3.1.

LLVM 21 compatibility fixes included here:

• rustc_llvm_wrapper: guard the llvm/Transforms/Instrumentation.h include to LLVM < 19 (the header is gone in LLVM 21; only the legacy pass registry uses it) and update the wrapper to the LLVM 21 APIs. tests/llvm-wrapper/ adds a small C++ check of the attribute/capture semantics the wrapper relies on.
• nvvm: restore kernel annotations after the LLVM 21 bitcode upgrade.
• cust: handle CUDA 13.3's new CU_GRAPH_NODE_TYPE_RESERVED_16 graph node type, which made an existing match non-exhaustive.

Extend the existing Linux CI matrix with LLVM 21 vector-add host/PTX compilation on x86_64 and aarch64, using CUDA 13.3.1 and LLVM 21.1.8 prebuilts. The six existing LLVM 7 configurations retain their workspace checks. There is no separate experiment workflow.

LLVM 21 prebuilts

The backend's default LLVM 21 download URL is the upstream Rust-GPU/rustc_codegen_nvvm-llvm llvm-21.1.8 release, which does not exist yet, so default downloads fail until it is published. That needs Rust-GPU/rustc_codegen_nvvm-llvm#4 merged and a release with linux-x86_64.tar.xz, linux-aarch64.tar.xz, and windows-x86_64.tar.xz.

Until then, the archive directory is overridable through PREBUILT_LLVM_URL (environment variable for the backend, repository Actions variable for CI; documented in the guide). The CI LLVM 21 jobs currently default to the same three archives published at https://github.com/brandonros/rustc_codegen_nvvm-llvm/releases/tag/llvm-21.1.8. Before merge, that default and the guide example should point at the upstream release.

Status

All LLVM 7 Linux jobs, both Windows jobs, and all container image builds pass, including the new Ubuntu-24.04/CUDA-13.3.1/LLVM-21.1.8 image.

The two LLVM 21 jobs compiled the wrapper and rustc_codegen_nvvm, then failed linking librustc_codegen_nvvm.so with unable to find library -lzstd: the prebuilts are built with zstd support, so llvm-config --system-libs emits -lzstd, and the job's container did not install libzstd-dev. The fix is to install libzstd-dev in that job.

Not ready to merge: pending the upstream release, a green LLVM 21 CI run, and runtime validation. The LLVM 21 jobs only compile the vector-add example; GPU execution and broader LLVM 21 workspace checks are not covered yet.

Based directly on upstream main; no portable PTX exporter changes are included.

[LegNeato]

 = note: /usr/bin/ld: cannot find -lzstd: No such file or directory
            collect2: error: ld returned 1 exit status
            

Looks like we need to install something?

[brandonros]

 = note: /usr/bin/ld: cannot find -lzstd: No such file or directory
            collect2: error: ld returned 1 exit status
            

Looks like we need to install something?

[brandonros]

@LegNeato Rust-GPU/rustc_codegen_nvvm-llvm#4 and then this whenever you're ready/if you have any feedback

thanks

[LegNeato]

Will merge once the other one builds, thanks!