nx_f32_conv2d_backward_gate.nx

nx_f32_conv2d_backward_gate.nx -- EXACT gradient check of nx_f32_conv2d_backward. Integer-valued inputs/weights make every gradient an exact small integer, so the analytic backward is checked BIT-EXACT (nx_i32_to_f32 canonical f32, ==) with NO numerical tolerance -- a stronger proof than finite-difference. Toy conv: C_in=1 H=3 W=3 input=[1..9], C_out=1 KH=KW=2 stride=1 pad=0 (OH=OW=2), weight=[10,20,30,40]. dOut=all-ones -> dW=[12,16,24,28], dBias=4, dIn(0,0)=10, dIn(1,1 center)=100 (verified by hand) dOut=[1,2,3,4] -> dW[0,0]=37, dBias=10 (general dOut weighting) stride=0 -> ERR_BAD_ARGS (guard) expect_exit: 0