Some fixes for MaternKernel

[devmotion]

This PR removes the loggamma definition and adjoint, removes the custom adjoint for kappa(::MaternKernel, ::Real), and fixes a type instability of this function. The loggamma function is already defined in SpecialFunctions and the custom adjoint is not needed when one uses ifelse instead of _ ? _ : _, as suggested in #114 (comment):

julia> using Zygote

julia> Zygote.gradient(0.0) do x
        iszero(x) ? one(x) : x^2
       end
(nothing,)

julia> Zygote.gradient(1.0) do x
        iszero(x) ? one(x) : x^2
       end
(2.0,)

julia> Zygote.gradient(0.0) do x
        ifelse(iszero(x), one(x), x^2)
       end
(0.0,)

julia> Zygote.gradient(1.0) do x
        ifelse(iszero(x), one(x), x^2)
       end
(2.0,)

Moreover, the PR fixes a type instability of kappa(::MaternKernel, ::Real). Currently

julia> @code_warntype KernelFunctions.kappa(MaternKernel(ν = 3.0), 1)
Variables
  #self#::Core.Compiler.Const(KernelFunctions.kappa, false)
  κ::MaternKernel{Float64}
  d::Int64
  ν::Float64

Body::Union{Float64, Int64}
1 ─      nothing
│   %2 = Base.getproperty(κ, :ν)::Array{Float64,1}
│        (ν = KernelFunctions.first(%2))
│   %4 = KernelFunctions.iszero(d)::Bool
│   %5 = KernelFunctions.one(d)::Core.Compiler.Const(1, false)
│   %6 = KernelFunctions._matern(ν, d)::Float64
│   %7 = KernelFunctions.ifelse(%4, %5, %6)::Union{Float64, Int64}
└──      return %7

whereas with this PR

julia> @code_warntype KernelFunctions.kappa(MaternKernel(ν = 3.0), 1)
Variables
  #self#::Core.Compiler.Const(KernelFunctions.kappa, false)
  κ::MaternKernel{Float64}
  d::Int64
  result::Float64

Body::Float64
1 ─      nothing
│   %2 = Base.getproperty(κ, :ν)::Array{Float64,1}
│   %3 = KernelFunctions.first(%2)::Float64
│        (result = KernelFunctions._matern(%3, d))
│   %5 = KernelFunctions.iszero(d)::Bool
│   %6 = KernelFunctions.one(result)::Core.Compiler.Const(1.0, false)
│   %7 = KernelFunctions.ifelse(%5, %6, result)::Float64
└──      return %7

[devmotion]

The test failures are very unexpected (but unrelated) - why would we test AD with negative values for the sigma^2 in ScaledKernel? And why was that not discovered when these tests were added? Maybe we have to set a seed somewhere?

[theogf]

Oh you're right I did not use the exp. transform here

KernelFunctions.jl/test/kernels/scaledkernel.jl

Line 43 in 9cfbdb5

test_ADs(x->x[1] * SqExponentialKernel(), rand(1))

I will make a very short PR

[devmotion]

I rebased the PR.