Add an opt-in first-party CUDA proving backend

[arthurpaulino]

Accelerate the Goldilocks/BLAKE3 prover with first-party CUDA while keeping the default CPU build independent of nvcc, the CUDA runtime, and NVIDIA hardware. The GPU path keeps DFT/LDE, mixed-height BLAKE3 commitments, lookup construction, quotient evaluation, batched openings, and binary FRI folding device-resident.

Preserve the protocol, proof format, proof sizes, transcript order, and CPU verification behavior. Canonicalize lazy Goldilocks representatives at proof serialization so CPU and CUDA proofs are byte-identical. This intentionally changes newly generated proof bytes relative to c72d321; old proofs remain verifiable, and README.md documents the compatibility boundary.

Harden the backend after adversarial review: keep pinned traces alive through CUDA handle destruction, reject unsupported Merkle caps and multi-bit GPU FRI, remove the racy generic FRI fold, canonicalize host uploads with grid-stride coverage, place kernel-visible metadata in managed allocations, derive spill headroom from device memory while reserving later-stage scratch, honor device selection, gate dynamic shared memory on device limits, remove the fixed constant-cache ceiling, validate FFI dimensions including one-row quotient steps, and discover nvcc through CUDA_HOME/CUDA_PATH.

Generate quotient-coset selectors directly in device memory from compact geometric parameters, using chunked recurrences and in-place batch inversion. This removes the largest remaining happy-path host-to-device upload without increasing peak VRAM. At Ix Vector.extract_append q50 scale, combined inner/outer quotient time falls from 1.73s to 1.45s.

Build native cubins for a supported common architecture set, add CUDA compile CI, a threshold-crossing CPU/GPU byte-compatibility smoke test, honest transfer-inclusive microbenchmarks, Plonky3 provenance, and dated benchmark documentation. Organize the Rust backend under src/cuda. The CUDA configuration currently supports Linux x86_64, cap_height=0, and binary FRI.

On an RTX PRO 6000 Blackwell, Ix Vector.extract_append recursive q50 reduces the post-execution-and-trace portion of inner and outer proving from 71.87s on CPU to 8.74s with CUDA (8.22x), with identical proof sizes and successful verification. End-to-end proving also includes Aiur execution and trace construction, which remain outside this comparison. Fifty queries is a development benchmark parameter, not a production security recommendation.