lengthscale transform

[st--]

There is currently ScaleTransform, ARDTransform, and LinearTransform. These could all be unified under a LinearTransform with potentially specialised matrices (e.g. a scaled identity matrix or a diagonal matrix for ScaleTransform or ARDTransform, respectively).

Moreover, other kernel/GP packages generally parametrise the kernels with the rather more interpretable lengthscale rather than the inverse lengthscale as KernelFunctions.jl does at the moment: it would be good to have an easy way for users to directly specify a lengthscale (ARD or not).

[devmotion]

I think the ScaleTransform is sufficiently different from the others since it does not operate on vectors necessarily.

Regarding ARDTransform, this was discussed in #107.

Length scale was discussed in #217 (comment), it seems the issue shouldn't have been closed since this was not resolved/implemented.

[devmotion]

The lengthscale issue is fixed by #336.