nx_train_int_vs_softfloat_gate.nx

nx_train_int_vs_softfloat_gate.nx -- THE PAYOFF measurement: does INTEGER training actually run FASTER than our SOFTWARE-FLOAT tape on IDENTICAL work? Trains the SAME 2-layer MLP (W1->ReLU->W2, L=sum(y-t)^2) for the same number of steps, TWO ways -- (A) software-float via nx_f32_mul/add (what nx_f32_qwen2_train uses), (B) integer Q16 -- times both with sys_now_us, and reports the speedup. Same algorithm, same MAC count; only the arithmetic differs, so the ratio IS the software-float tax paid in a real training loop. This is the internal h2h that justifies porting the trainer to integer (the sovereign path to beat PyTorch CPU). Measured-not-asserted. ReLU + sign checks read the IEEE sign bit directly ((raw>>31)&1) to avoid nx_int-typed comparisons. license_tier: ORIGINAL