Complete restructure

[milankl]

@avleenk2312 @tomkimpson This PR will completely restructure this package because we had badly defined things too specific and so it does this now as general.jl which defines all functions for ::AbstractStochasticFloat and types.jl which defines only what's necessary for a specific type, e.g. BFloat16sr is just

primitive type BFloat16sr <: AbstractStochasticFloat 16 end
Base.float(::Type{BFloat16sr}) = BFloat16       # corresponding deterministic float
stochastic_float(::Type{BFloat16}) = BFloat16sr # and stochastic float
BFloat16sr(x::BFloat16) = stochastic_float(x)   # direct conversion
Base.uinttype(::Type{BFloat16sr}) = UInt16      # corresponding uint
Base.widen(::Type{BFloat16sr}) = Float32        # higher precision format for exact arithmetic

"""Stochastically round x::Float32 to BFloat16 with distance-proportional probabilities."""
function stochastic_round(T::Type{BFloat16},x::Float32)
    ui = reinterpret(UInt32,x)
    ui += rand(Xor128[],UInt16)			        # add stochastic perturbation in [0,u)
    return reinterpret(BFloat16,(ui >> 16) % UInt16)	# round to zero and convert to BFloat16
end

As you see this frees a lot of lines of code that were previously used to redefine for every type which could be just collected into a single more general formulation. Draft for now as I still need to adapt all tests but it also DONE.

fixes #72 #70 #63

@madnanabid

[milankl]

This will be v0.8 because I've removed BFloat16_stochastic_round etc in favour for a more general stochastic_round(T::Type{<:AbstractFloat},x)