feat: Add SGD, Momentum, and Adam with same-net comparison

[ThomasHartDev]

Adds the three standard first-order optimizers as pure numpy update rules that sit next to the hand-written MLP backprop. Each class owns its own state (velocity for momentum, bias-corrected moments for Adam) and steps a flat list of parameter arrays in place, so the training loop can swap rules without touching the gradient math.

compare_optimizers retrains the same architecture on the same data with the same seed for each factory, which keeps init and minibatch order fixed. On XOR all three cut loss; Adam usually pulls ahead early because the adaptive rates absorb uneven gradient scale. Defaults match the common library settings so the curves are easy to reason about against PyTorch.