Added Multi Scale Fourier Feature Network as per Issue SciML/NeuralPDE.jl#498

[Parvfect]

Considering the amount of code rewritten, it makes more sense to define it as a struct or purely define one ff layer that encapsulates a fourier feature encoding and dense layers and then have a different file where users can create msff networks using FastChains or a new file for Networks that are defined using general FastChains.

Addtionally, bias is not included because of my lack of understanding of including it in a solitary function. It has not been rigirously tested, although it seems to compile. I did get an inexact error that I couldn't debug when trying to run some operations, I would need some help to tackle the issue now that I know the general way that the network can be implemented.

Generally don't think that this network belongs in the fastlayers.jl file but need someone to guide me where to put it.

[ChrisRackauckas]

Needs tests. Given that MSFF isn't even defined, I don't think this has been tested as all.