fix(test): row-pair leaf layout in merkle_root_parity GPU parity tests

[MauroToscano]

What

Two GPU↔CPU root-parity cases in crypto/math-cuda/tests/merkle_root_parity.rs compare against the wrong Merkle leaf layout on main:

• gpu_and_cpu_row_major_merkle_roots_match
• gpu_and_cpu_ext3_merkle_roots_match

Their GPU helpers (gpu_merkle_root / gpu_ext3_merkle_root) call keccak_leaves_base / keccak_leaves_ext3 with rows_per_leaf = 1 (per-row → num_rows leaves), but the CPU reference commit_rows_bit_reversed uses the row-pair layout (ROWS_PER_LEAF = 2 → num_rows / 2 leaves, each hashing a bit-reversed row pair). Different leaf count and different leaf bytes ⇒ different root, so these two cases can never match main's row-pair commitment scheme.

Fix

Pass rows_per_leaf = 2 at both call sites so the generic GPU keccak-leaves + Merkle path uses the same bit-reversed row-pair layout as the CPU commit. (Plus updated the now-inaccurate "per-row" comments.)

Why this is a test fix, not a production bug

• Test-only surface. keccak_leaves_base / keccak_leaves_ext3 have no production callers — the proving path uses the fused coset_lde_*_row_major_with_merkle_tree_keep, which is already row-pair. GPU/CPU proof parity was never affected, which is exactly why proofs verify with and without the GPU.
• The row-pair primitive is already proven. tests/keccak_leaves.rs::keccak_leaves_base_row_pair_matches_cpu (and the ext3 equivalent) already verify that keccak_leaves_base(.., 2) matches the CPU row-pair prover helper.
• The production-path cases in this same file already pass. new_row_major_pipeline_base_root_matches_cpu / …_ext3_… exercise the real fused pipeline (row-pair on both sides) and are green.

Background

These two cases were introduced in #715 (per-row GPU leaves). #735 ("unify & clean up the commitment layer") moved the scheme to row-pair — adding ROWS_PER_LEAF = 2, the row-pair leaf kernels, and a row-pair commit_rows_bit_reversed — but left these two test calls at the old rows_per_leaf = 1. This realigns them. The mismatch went unnoticed because GPU tests don't run in CI (no nvcc).

Testing

Not run locally — no GPU/CUDA on the dev box, and math-cuda only builds under the cuda feature. To be confirmed on a GPU machine: all four cases in merkle_root_parity.rs should pass. rustfmt clean.