Logistic regression + cross-entropy loss

[ThomasHartDev]

adds binary logistic regression trained by minibatch SGD, in the same shape as the linear regression module already here. the interesting bit is that the gradient of the cross-entropy w.r.t. the logits is just sigmoid(z) - y, the same residual you get from least squares, so the training loop looks almost identical. the sigmoid is sign-branched so exp never overflows and the loss uses softplus(z) - y*z via logaddexp so a confidently-wrong prediction gives a big finite number instead of inf, which is the usual place a naive log-loss falls over. also threw in a decision_boundary helper that gives you the 50 percent line for a 2-feature model so you can actually plot the thing. tests cover the numerical-stability edges, separable data hitting 100 percent, l2 shrinkage, and the usual bad-input guards anyway.