feat(nn): Linear layer and activation functions

[Celz-Pch]

Goal

First building blocks for neural networks.

Tasks

• nn::Linear(in_features, out_features) with weight init
• Activations: relu, sigmoid, tanh
• nn::Module base class with forward() and parameter registration

Acceptance criteria

• Can build a 2-layer MLP and run a forward pass
• Parameters are discoverable for the optimizer

Depends on

• autograd issue