GPU offload of a decode-heavy expansion: check GPU busy % and the CPU baseline's effective cores before claiming a speedup

A GPU path for expanding compressed per-signal change records looked 1.89x faster than the CPU path in wall time, but the kernel was 3.3% of wall and the GPU averaged ~5% busy; ~95% of the GPU path's time was single-threaded host-side decode/packing. Separately, the CPU baseline was only using ~1.3 of 4 cores (CPU-s / wall), due to the parent draining worker pipes one at a time. Before reporting a GPU win: (1) report GPU busy % and kernel+transfer time vs total wall, (2) compute CPU-s/wall for the CPU baseline to get its effective core count, (3) parallelise the host prep and fix the CPU baseline, then compare. When the per-element work is cheap relative to decode, a fixed multi-process CPU path usually matches the GPU path without the extra hardware cost.