docs(relax): §15 gridded SSD MR scan re-profile — 8L ≈894 tok/s / 28L ≈293 tok/s, gap 75×→≈3.8× / 289×→≈12×, bottleneck flips to abstract numpy bwd

Re-measured gridded CUDA SSD chunked scan live on gb10 sm_121 (probe_chunked_scan_cuda_gb10.py
--prod): F2 0.980 ms vs serial 6.56 s ≈ 6694×, fp16 parity 4.746e-04 PASS, grid 28672 tg;
F0 5.025 ms, F1 1.226 ms; delegation builds compiled CUDA JITKernels at prod shape (PASS).
Substituted measured gridded fwd per-call (7.231 ms) for serial MR (6559.7 ms) into the §14
step (labelled extrapolation, self-checks to §14 45.44 tok/s within 0.4%). No memory cost
(planned peak unchanged). Forward collapses 95.0%→~2% of step; abstract numpy backward becomes
the dominant term (79.6% @8l / 91.4% @28l) — next lever is device-resident + gridded-SSD bwd.

Author: apstenku123