fix: admit scalar arithmetic on verified pre-Ampere targets

[ssoj13]

The generator hardcoded PTX 7.0/sm_80 for every scalar-arithmetic variant and required evidence to match those constants. This rejected valid pre-Ampere mul_rn_f32/fma_rn_f32 calls.

Separate the conservative native family requirement (PTX 2.0/sm_20) from the independently verified LLVM route (PTX 3.2/sm_20) and CUDA 13.3 libNVVM route (PTX 6.3/sm_75). Refresh evidence and regenerate the catalog, target tables, wrappers, ABI documentation and book stamp. The compiler's existing target vocabulary is unchanged.

Port of ansidium/cuda-rust-windows#14, independently rebased onto NVIDIA main bc46b1ffb and regenerated with its pinned inputs on Linux. Related: #1462. This covers all 64 scalar_arithmetic variants; scalar_math is outside this change, so it does not close the whole issue.

Verification:

• All generator/codegen tests, current-output check, ABI-history check against NVIDIA main, clippy/all targets with -D warnings, formatting, Ruff, SPDX and book stamp checks pass.
• Committed reproducible captures cover 192 LLVM codegen/assembly cases and canonicalization of all 64 declarations, using the pinned LLVM 23.1.3 source.
• CUDA 13.3 evidence compiles/assembles all 64 operations for compute_75 and compiles/LTO-links retained kernels; both cubins contain all 64 entries. Fixture runtime execution is not claimed.
• Independent semantic audit confirms only target/backend fields of 64 records change; the other 965 records, all 1029 ABI identities, LLVM signatures and opcodes are unchanged. Most generated-file changes are catalog hash comments.

The change and current verification were prepared with Codex. Existing source attribution is retained; new probe helpers retain attribution to the repository contracts they adapt and identify the adaptation.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.