feat(polynomials): Accumulator<Fr> + vectorize compute_sum

[notnotraju]

Goal. Introduce `Accumulator` — the reduction counterpart to `Broadcast`. `operator+=(Fr)` for ScalarIndex contributions, `operator+=(VectorField)` for ContiguousVectorIndex contributions, `reduce()` horizontal-adds the N vector lanes with the scalar slot. Migrate `polynomial_arithmetic::compute_sum` as the canonical first user.

Non-goal. Migrate other reduction sites. Accumulator is the primitive that compose with Polynomial's existing operator[]; richer reduction kernels (sumcheck partial-evaluate, MLE fold, IPA inner product) live in follow-up work.

Why this is a separate PR. Accumulator is the third of the adapter types (after Broadcast and the implicit Polynomial proxies). It's the design building block for any kernel that reduces over per-iteration contributions, including the dot-product shape (`acc += a[ctx] * b[ctx]`) needed for IPA. Reviewable independently of the abstraction-layer PR — the diff is one new adapter type + one migrated function + one bench file.

Kernel after migration.
```cpp
PolynomialSpan view{0, std::span(src, n)};
Accumulator acc;
vectorized_for<VECTOR_FIELD_WIDTH>(0, n, [&](auto ctx) { acc += view[ctx]; });
return acc.reduce();
```

V8/WASM bench (N=2^16, n=5 reps). From barretenberg-claude where the code originated; file states are bit-identical:

• `compute_sum_scalar` median: 582 µs (CV 15%)
• `compute_sum_full` median: 357 µs (CV 1.7%)
• Speedup: 1.63×

WASM op-count diff (same bench binary, scalar vs vectorized source):

• `v128.and` (29-bit mask): 0 → 26
• `v128.or` (carry): 0 → 7
• No `extmul_*` — compute_sum is addition-only; SIMD primitives appearing are VectorField's lane-wise add kernel, not Mont-mul. Correct.

Tests. `polynomials_tests` passes everything mt/bb passes today. `CorrectnessGuard` at `compute_sum_bench` startup verifies bit-equality of scalar vs vectorized on 65k random fr's.

Stack.

• Depends on rk/wasm-simd-02-vectorized-for
• (No further dependents in this stack; future reduction migrations e.g. IPA inner product will build on this.)