Gives us a class to implement more activation functions.

Inherit from Layer, implement getOutputImpl().

We can then implement different optimizers that the user can choose.

I'm a bit inconsistent with my formatting; I oscillate between camel and snake case. Though there isn't a c++ standard for identifier names, lets shoot for camel case for function and variable names (though most is using camel case).

As the title implies, the framework currently iterates through each neuron in the layer to grab the weights, which is a performance hit. We can move the weights to Layer, but this essentially makes Neuron trivial. May keep Neuron in the future.

This way, we can add additional cost functions. Two public functions: cost(...), cost_prime(...) that determine the right functions based off of the std::string cost. Then a bunch of protected functions that implement different cost functions.

Add checks for parameters such as first layer input size = x.size(), output size = y.size(). Add tests to verify these checks.

Good milestone to show in the example code.