Added more activation function support

[grantbruer]

This PR lets users use more activation functions.

It has three main parts:

1. Activation functions are no longer restricted to be invertible or non-invertible for certain networks. The previous behavior would silently give incorrect gradients if an invertible function was given where a non-invertible function was expected and vice versa.
   • I implemented this by making a common interface apply_backward() for backpropagating through invertible and non-invertible functions.
2. I implemented softplus and hyperbolic activation functions.
3. I added tests for these changes.

I based it off slimgroup#119, which I expect to be accepted first.

[grantbruer]

Probably apply_backward() should be renamed to backward(). I was initially worried about some name clashing with that name, but it is consistent with all the other backward functions.