fix(bb): pad gemini_masking_poly size to even to avoid sumcheck OOB

[AztecBot]

Summary

Fixes the nightly debug-build failure of HonkRecursionConstraintTestWithoutPredicate/1.GenerateVKFromConstraints, which has been failing every night since 2026-04-22.

Root cause

Introduced by #22334 (top-of-trace masking), OinkProver::commit_to_masking_poly allocates gemini_masking_poly at max_end_index() instead of the dyadic size. For this test, max_end_index() = 4135 (odd).

Sumcheck's compute_effective_round_size rounds max_end_index up to the next even number (4136) so its pairwise (edge_idx, edge_idx+1) iteration is well-formed. This causes a read of gemini_masking_poly[4135], one past its allocation. In release builds SharedShiftedVirtualZeroesArray::get returns 0 for OOB and the spurious read is harmless. With -D_GLIBCXX_DEBUG and active BB_ASSERTs, the nightly debug preset trips BB_ASSERT(index < virtual_size_ + virtual_padding).

ECCVM and Translator allocate their masking polynomials at circuit_size (a power of two), so they were unaffected.

Fix

Match sumcheck's even rounding when allocating the masking polynomial:

const size_t max_end_index = prover_instance->polynomials.max_end_index();
const size_t polynomial_size = max_end_index + (max_end_index % 2);

This is the same change proposed in #22789 (auto-closed for staleness).

Verification

cd barretenberg/cpp && cmake --preset debug
cd build-debug && ninja dsl_tests
./bin/dsl_tests --gtest_filter='HonkRecursionConstraintTestWithoutPredicate/1.GenerateVKFromConstraints'
# [  PASSED  ] 1 test. (164778 ms)

Failing nightly run: https://github.com/AztecProtocol/aztec-claude/actions/runs/25204419411
Detailed analysis: https://gist.github.com/AztecBot/539dcf1fe5bf600319f08d19c883285d

ClaudeBox log: https://claudebox.work/s/22d70e3e00df70b6?run=1