bench(v4): regen gb10 op-level matrix — all 8 GDN/KDA path_b cells now ok on CUDA

Was: every path_b cell 'failed: [metal_kernel] No Metal back-end.'
Now: ok via TileLang-CUDA EAGER bridge, output matches path_a to ~7e-8,
median ~1.2-2.8 ms (vs path_a ~24-40 ms). SHA 3ef4894.

Author: apstenku123