A Psf cannot in general be used as a Kernel unless it is a KernelPsf for good reason: many of our Psfs are not continuous, and in general a Psf may be quite expensive to evaluate at a point.
We should have a Task that computes a Kernel that approximates a Psf by sampling it at discrete points and assuming some interpolation. This interpolation will be incorrect in general, but I imagine us only using this in contexts where approximation is acceptable (i.e. in smoothing an image for detection). I imagine a LinearCombinationKernel is probably the best approach here, possibly with an image PCA over the sampled Psf images used to define the basis images. The algorithm would thus look much like the PcaPsf determiner algorithm, but with no concern for outliers, blends, or noise, and it would be responsible for sampling the input PSF in a configurable way.
This would be very useful for a more complete version of DM-10325, and I believe David Reiss recently expressed a need for this functionality as well (though I'm not sure how much his use case can tolerate approximation).