Neural network implementation

[decahedron1]

Replace the library with a far more accurate (60% vs 95%), only slightly slower (1.3us vs 22us) neural network-based implementation. This implementation is still over 10x faster than competing neural network VADs Silero & TEN.

There's still much room for optimization, especially outside of the NN inference. #![no_std] is currently broken since we rely on std f32 math; these areas should use libm (or something else) instead.