Add i-times integrated Wiener Kernel

[sharanry]

Issue #44

i-times integrated Wiener process kernel function given by

    κ(x,y) =  kᵢ(x,y)

For i=-1, this is just the white noise covariance, see WhiteKernel.
For i= 0, this is the Wiener process covariance,
for i= 1, this is the integrated Wiener process covariance (velocity),
for i= 2, this is the twice-integrated Wiener process covariance (accel.),
for i= 3, this is the thrice-integrated Wiener process covariance. where kᵢ is given by

    k₋₁(x,y) =  δ(x,y)
    i >= 0, kᵢ(x,y) = 1/ai * min(x,y)^(2i + 1) + bi * min(x,y)^(i+1) * |x-y| * ri(x,y),
    with the coefficients ai, bi and the residual ri(x,y) defined as follows:
        i = 0, ai =   1, bi = 0
        i = 1, ai =   3, bi = 1/  2, ri(x,y) = 1
        i = 2, ai =  20, bi = 1/ 12, ri(x,y) = x + y - 1/2 * min(x,y)
        i = 3, ai = 252, bi = 1/720, ri(x,y) = 5 * max(x,y)² + 2xz + 3min(x,y)²

References:

• See the paper Probabilistic ODE Solvers with Runge-Kutta Means by Schober, Duvenaud and Hennig, NIPS, 2014, for more details.
• http://www.gaussianprocess.org/gpml/code/matlab/doc/manual.pdf pg44
• https://github.com/alshedivat/gpml/blob/master/cov/covW.m

[theogf]

Maybe that would make sense to make separate kernels for each i, like for the Matern Kernel

[sharanry]

@theogf Considering we are using parameterized types right now, would separate kernels make much of a difference in terms of performance? Currently, IMO even the usage seems straight forward.

[willtebbutt]

Looking good so far. If you could sort out style as in the other PR, that would be great.

[sharanry]

@willtebbutt Do you think this is ready to be merged?

[sharanry]

@devmotion @theogf @willtebbutt Can this be merged?

[willtebbutt]

Other than the style thing, I'm happy for this to be merged.