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. The modern Nix shell uses LLVM 21.1.8 and CUDA 13.3, and CUDA 13 Docker images use the published CUDA 13.3.1 bases.

Update feature forwarding, LLVM tool discovery, CI, the vector-add example, and documentation together. Modern builds require an installed LLVM 21 toolchain or an explicit PREBUILT_LLVM_URL until LLVM 21 prebuilts are published. Update the Windows CUDA installer action to support CUDA 13.3.1.

Add a Linux x86_64/aarch64 experiment workflow that uses the successful LLVM 21.1.8 artifacts from prebuilt run 34781677658 to build the existing vector-add host and PTX with CUDA 13.3.1. This temporarily pins Actions artifacts, which expire; replace them with release assets when published. The workflow requires no GPU and does not execute the kernel.

This experiment is based directly on main and contains no portable PTX export changes.

Validation: diff checks, workflow YAML parsing, and embedded shell syntax checks pass. Both Linux prebuilt archives were downloaded and their tool layout checked. Rust-CUDA compilation with those artifacts is pending; LLVM wrapper API changes may still be needed. GPU runtime validation and local container-image builds have not been performed.

[brandonros]

Superseded by the upstream draft PR: Rust-GPU#408. The experiment branch is retained; further review and integration will happen there.