Add portable PTX export and verified modern LLVM cleanup

[brandonros]

Depends on #408 (LLVM 21.1.8 / CUDA 13.3). This draft targets upstream main, so its current diff includes that prerequisite. After #408 merges, rebase this branch onto upstream main to leave only the export and optimization changes.

Add a GPU-independent Rust-to-PTX exporter, restore default merged-module GlobalDCE on the modern backend, and expose optional verified cleanup pipelines. Include retention regressions, modern NVVM shuffle wrappers, guarded-select fixtures, and tools for inspecting PTX/SASS and comparing mining workloads.

The exporter, feature gates, intrinsic assembly, scripts, and CI caches use LLVM 21. Existing experimental Llvm19Cleanup / llvm19_* API names remain for caller compatibility. Checked-in LLVM 19 measurements remain historical evidence; they do not establish LLVM 21 correctness or runtime performance.

The native stack is maintained in the fork: brandonros#3 → brandonros#4. The top fork PR provides the incremental diff above the LLVM upgrade. Both upstream PRs use these same branch heads.

Validation: all nine Python unit tests, offline Cargo metadata with ptx_export/llvm21, touched Rust formatting checks, and whitespace checks pass. Full Linux LLVM 21/CUDA compilation and GPU numerical validation of the combined branch remain pending. Optional scalar cleanup and inlining remain disabled by default.