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.

Extend the existing Linux CI matrix with LLVM 21 vector-add host/PTX compilation on x86_64 and aarch64. These jobs use CUDA 13.3.1 and LLVM 21.1.8 prebuilts from Rust-GPU/rustc_codegen_nvvm-llvm. The backend's default LLVM 21 download uses the same upstream release. The six existing LLVM 7 configurations retain their workspace checks. There is no separate experiment workflow or dependency on contributor-hosted releases or Actions artifacts.

Depends on Rust-GPU/rustc_codegen_nvvm-llvm#4 and publication of the upstream llvm-21.1.8 release with linux-x86_64.tar.xz, linux-aarch64.tar.xz, and windows-x86_64.tar.xz. The URLs intentionally describe the post-publication setup; downloads will fail until upstream publishes that release.

Validation: Rust formatting, workflow YAML parsing, and diff checks pass. Trial LLVM 21 prebuilts built successfully for all three platforms. Rust-CUDA integration exposed unresolved compatibility failures: the wrapper includes llvm/Transforms/Instrumentation.h, absent in LLVM 21, and cust has a non-exhaustive match for CUDA's CU_GRAPH_NODE_TYPE_RESERVED_16. This remains a draft pending the upstream release, compilation fixes, and runtime validation. The LLVM 21 CI jobs compile the existing 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.