Adding activation functions and fast LUT implementation

[jfsantos]

This PR adds the SiLU (Swish) and HardSwish activations to core, as well as changes LeakyReLU into a parameterizeable function. It also adds an activation class that converts any C++ function with one float input argument into a LUT implementation with N points and linear interpolation between points.

Tests were added for the fast LUT implementation to make sure the results are close to the standard C++ implementations. With a LUT using 1024 entries, we get the following results in the tests for sigmoid(1.25) and tanh(1.25):

exact sigmoid: 0.7773
lut sigmoid:   0.777297
exact tanh: 0.848284
lut tanh:   0.848269

The LUT implementations can be activated to replace the original activations just like it is currently done for fasttanh.

Developed with support and sponsorship from TONE3000

[sdatkinson]

LGTM 👍🏻