TypedPython: vectorise element-wise loops over array.array buffers without giving up per-element atomicity #226

Follow-up to #149/#225 (leader decision 2026-10-17: keep relaxed atomic element access; plain access is a data race in the C memory model, which N-9 does not allow).

Measured (quiet window, 11 interleaved pairs, 2,000,000 doubles, 3.15.0rc1t):

• an element-wise store loop (a[i] = a[i] * k) is 2.54x slower with relaxed atomics than with plain access, because the compiler does not vectorise atomic accesses;
• a read-and-reduce loop is unaffected (1.001), since the strict-FP reduction is not vectorised anyway.

Idea to investigate: emit explicitly vectorised code whose vector loads and stores are element-wise single-copy atomic on the target (aligned NEON / SSE / AVX element accesses are single-copy atomic per element on arm64 and x86-64). Express this so the C memory model still sees per-element atomic accesses, for example a target-specific helper with a documented guarantee. Prove or measure that no tearing below 8 bytes can occur.

Done when: the scale kernel's cost is within 10% of plain access, and a design note states per target why each element access stays single-copy atomic.