TypedPython: re-measure #226 vector buffer loops in a quiet window; revert if not within 10% of plain #274

PR #270 (#226) measured 0.94x of plain access only once, at load 8.7. Re-measure the scale kernel (plain, relaxed-atomic scalar, inline-asm vector; benchmarks/typedpython/results/2026-10-17-buffer-atomic-kernels.py) in a quiet window (load below 1, 11 interleaved pairs).

Done when: the quiet-window numbers are recorded; if the vector path is not within 10% of plain (or not faster than the scalar atomic path), the vector path is reverted (cgen loop matcher and tp_buf_ld2_f64/tp_buf_st2_f64) with the design note updated.