feat: Add activation functions and Xavier/He weight init

[ThomasHartDev]

Pulled the nonlinearities and the init schemes into their own module so the variance story is measurable on its own. Each activation carries a local backward for act'(z), and Xavier vs He are the usual closed forms (2 over fan_in+fan_out for symmetric acts, 2 over fan_in for ReLU). The interesting bit is forward_variance_profile: stack a few layers from unit noise and you can watch naive N(0,1) weights explode, He hold a ReLU stack near O(1), and Xavier quietly starve the same ReLU stack. That last case is the silent failure mode I wanted to pin down with a test rather than leave as folklore.

Derivatives are checked against finite differences, sample variances of the draws match the formulas, and the MLP path is left alone for now so this stays a focused slice.