feat(cuda): optional arch-tiered PTX dispatch with compute_61 fallback

[jamesburton]

Implements the minimal first step from #332: an optional, additive arch-tiered CUDA kernel dispatch mechanism, with compute_61 portable PTX retained as the universal default and fallback.

Zero behavioral change by default. No higher-arch PTX is shipped, so the loader resolves to exactly the same files as today on every GPU. This PR is the extension point only — no new/optimized kernels — so the mechanism can be reviewed in isolation. Arch-specific kernels (e.g. tensor-core prefill GEMM) would be follow-ups once the approach is agreed.

What it does

• CudaModule.ResolveArchVariantPath / LoadForArch: for a kernel, prefer the highest <kernel>.sm_<arch>.ptx whose arch ≤ the device compute capability; otherwise use the universal <kernel>.ptx. Pure, allocation-light file-system logic.
• CudaKernels(string ptxDir, int ccMajor = 0, int ccMinor = 0): backward-compatible overload; the 0 default selects the universal fallback (today's behavior). CudaTransformerModel and HybridTransformerModel pass the already-detected CudaDevice.ComputeCapabilityMajor/Minor through.
• native/build.sh / build.ps1: opt-in higher-arch emission via EXTRA_ARCHS (e.g. EXTRA_ARCHS="80 86"). The default invocation produces only the compute_61 PTX — the curated kernel list is empty until a genuinely arch-specific kernel exists.

Why

compute_61 is a virtual arch, so the shipped PTX cannot encode sm_75+ ISA (tensor-core mma.sync, cp.async, bf16) regardless of the physical GPU. This adds a per-kernel opt-in path to ship higher-arch variants while preserving the single-portable-PTX deployment model and full Pascal+ back-compat. Rationale, the fatbin alternative, and the recommendation are in #332.

Validation

• dotnet build -c Release — clean (0 warnings).
• New CudaArchVariantSelectionTests (7 cases: base fallback, matching/lower arch, highest-eligible, CC 0, malformed token, no cross-kernel leakage) — pass.
• End-to-end GPU smoke (RTX 3060, sm_86): a Q4_K_M model loads via --device gpu and generates correctly, confirming the arch-aware loader is behavior-preserving on the universal fallback path.

Marked draft — happy to adjust scope/naming or fold in a first arch-specific kernel if you'd prefer the mechanism land alongside a concrete user. Refs #332.