feat(cuda): R4 DEEP composition + FRI commit phase on GPU

[ColoCarletti]

Summary

Extends the GPU-resident proving pipeline through Round 4. R4 DEEP composition and the full FRI commit phase (fold + per-layer Keccak leaves + pair-hash Merkle
tree) now run device-side, with only per-layer roots D2H'd for the transcript. The R2 composition-parts LDE moves to a _keep variant so its de-interleaved device
buffer is retained on Round2 and reused by R4 DEEP without a re-H2D. Also lands a Blelloch chunk-scan parallel batch-inverse kernel as infrastructure for future
GPU-side denominator inversion (not yet wired).

Changes

• crypto/math-cuda/kernels/{inverse,deep,fri}.cu — new kernels.
• crypto/math-cuda/src/{inverse,deep,fri}.rs — host orchestrators including FriCommitState (ping-pong eval buffers, in-place inv_twiddles squaring, per-layer
  fused fold + leaves + tree).
• crypto/stark/src/gpu_lde.rs — new dispatches: try_evaluate_parts_on_lde_gpu_keep, try_deep_composition_gpu, try_fri_commit_gpu. New counters:
  gpu_deep_calls, gpu_fri_calls.
• crypto/stark/src/prover.rs — Round2.gpu_composition_parts holds the R2 keep handle; R4 DEEP fast path inside compute_deep_composition_poly_evaluations
  consumes R1 main/aux + R2 parts handles when available.
• crypto/stark/src/fri/mod.rs — commit_phase_from_evaluations routes through try_fri_commit_gpu when cuda is enabled.
• Tests: parity for batch invert (n in {2..2^20}), DEEP, FRI per-layer tree (log_num_leaves in {1..18}); cuda_path_integration asserts the two new counters fire
  end-to-end.

Fallback

Every dispatch is gated by TypeId checks (Goldilocks + ext3) and the LDE-size threshold. Below threshold or on any cudarc error, the dispatch returns None and
the existing CPU implementation runs unchanged. Exception: mid-FRI-loop cudarc failure panics, because the transcript is already advanced and a CPU restart would
re-sample zeta_0 against mutated state.

Test plan

• cargo test -p math-cuda --release --tests (GPU host) — 67 tests
• cargo test -p stark --release --features cuda — 128 tests
• cargo test -p stark --release (no cuda) — 128 tests
• cargo test -p lambda-vm-prover --release (no cuda) — 384 tests
• cargo test -p lambda-vm-prover --release --features cuda --test cuda_path_integration -- --ignored — all 6 counters fire end-to-end
• cargo clippy --workspace --all-targets --features cuda -- -D warnings — clean
• cargo clippy --workspace --all-targets -- -D warnings — clean
• cargo fmt --all --check — clean