nx_f32_conv_train_gate.nx

nx_f32_conv_train_gate.nx -- proves the SOVEREIGN conv TRAINING LOOP converges end-to-end: forward (nx_f32_conv2d_forward) -> MSE loss/grad (nx_f32_train_ops) -> backward (nx_f32_conv2d_backward) -> SGD step, iterated. A 2x2 conv (C_in=1 H=2 W=2, C_out=1, 1x1 out) is trained from ZERO weights to fit a target produced by a known "true" weight; the loss must never increase and must converge to ~0 -- the first mechanical evidence that "build 2" (training our own pose net) works, reducing the arc to SCALE + real labeled data. Also unit-checks the ReLU backward and the MSE gradient (needed for the multi-layer pose net). This is a CONVERGENCE proof, NOT a trained pose net. expect_exit: 0