nx_q4k_speed_bench.nx

nx_q4k_speed_bench.nx -- measure the CPU speedup of the integer fused-dequant-dot vs emulated-f32. sd-server -> Nishi migration (CPU perf payoff). Both paths now agree with ggml (dequant tech debt fixed). This times, on the real blk.0.attn_q Q4_K weight row (4096), N iterations of: (A) INTEGER: nx_q4k_dot_row_col (fused dequant + integer-Q34 dot; the SOTA quantized-GEMM route). (B) F32: nx_q4k_to_f32 (dequant to f32) + emulated-f32 dot (the correctness route). Activation is converted to f32 ONCE outside the timed loop (amortized in a real GEMM), so we compare the per-output-neuron cost fairly. Reports us/op each and the speedup x100. license_tier: ORIGINAL