fix(math-cuda): driver-independent cuda builds (cudarc pin + cubin AOT)

[MauroToscano]

Visual explainer (the GPU residency work at a glance — what changed, what existed before, and the numbers): https://gistpreview.github.io/?a6bc0786b46cefe50b330ccb4d8f77e9

What

Makes cuda builds driver-independent so a box whose driver is older than the build toolkit still runs on the GPU instead of silently falling back to CPU (or panicking at init). Two independent robustness fixes plus bench-script cleanup.

The GPU constraint-eval + trace-residency work this supported has already merged (#798, #799); this is the build-robustness follow-on, now standalone against main.

The two fixes

1. AOT cubin instead of PTX. build.rs now compiles each kernel to a native cubin (SASS) for the host GPU's real arch (nvcc --cubin -arch sm_XX, loaded via Ptx::from_binary) instead of emitting PTX for the driver to JIT. A toolkit's PTX ISA version is fixed by its CUDA version, and a driver older than that toolkit rejects the module at load with CUDA_ERROR_UNSUPPORTED_PTX_VERSION → every kernel silently falls back to CPU. A cubin carries pre-compiled SASS, so the driver loads it directly regardless of toolkit version. Trade-off: a cubin is arch-specific — we build and run on the same GPU box in every flow and detect the arch from nvidia-smi, so this is exactly right; CUDARC_NVCC_ARCH overrides for cross-compiling.

2. Pin cudarc to cuda-12080. Replaces cuda-version-from-build-system + fallback-latest, which bind the newest symbol set the build toolkit knows (e.g. a 13.1 toolkit pulls in cuDevSmResourceSplit); cudarc resolves those eagerly at init, so a driver that predates them panics. Our cudarc surface is entirely CUDA-11-era, so the 12.8 set is a strict subset every ≥12.8 driver exports. This replaces the fragile per-script sed pin.

Driver floor: the pin requires a driver of CUDA ≥ 12.8 (branch 570+; any Blackwell-capable driver qualifies). On an older driver cudarc aborts at init rather than CPU-falling-back — documented in the README.

Also in this PR

• Fail-loud instead of silent-CPU. Warn-once at runtime when the GPU backend is unavailable; and at build time, when nvcc is present but no GPU arch can be detected and no CUDARC_NVCC_ARCH is set, the build hard-fails rather than guessing. A cubin is arch-locked, so a guessed arch (the old sm_89 fallback) would load on exactly one GPU model and silently CPU-fall-back on every other card — including lower ones like the 3090 (sm_86), which can't even run an sm_89 cubin. A hard error is loud and fixable (set CUDARC_NVCC_ARCH or build on the target host); a guess is neither. (nvcc-absent hosts take the empty-stub path and never reach this, so CPU-only CI is unaffected.)

The ABBA bench-script cleanup that originally rode along here is split into #812 — trivial, no build or runtime impact.

Validation

Validated on RTX 5090 boxes with 580 and 590 drivers (13.1 toolkit) — exactly the driver-older-than-toolkit case where the previous PTX path would have failed to load; the cubin path runs with default env, no CUDARC_CUDA_VERSION or CUDA_HOME juggling. make lint green on both the default and cuda feature sets.