Single-precision >= boolean mask (no vDSP equivalent).

Computes $\text{out}_i = (a_i \ge b_i) \,?\, 1 : 0$ via a scalar loop; used by autograd for ReLU/Max gradient masks where the comparison result must be a float.