nx_f32_train_ops.nx

nx_f32_train_ops.nx -- the small reusable software-f32 TRAINING ops that close the conv training loop, composing the gated f32 tower (nx_f32_mul/add/sub/gt) with the new nx_f32_conv2d_backward. These are the shared primitives the sovereign pose-net training (operator: "build 2") uses every step: MSE loss + its gradient (heatmap regression is MSE over the target heatmaps), the ReLU backward (gradient flows only where the activation fired), and an SGD parameter step. No optimizer state here (Adam lives in nx_f32_adam); this is the minimal, gate-able core. license_tier: ORIGINAL