Improve PSF modeling on L3 images

[stscijgbot-rstdms]

Issue RCAL-982 was created on JIRA by Jonathan Eisenhamer:

From github#⁠1495:

We compute PSF fluxes and positions on both L2 and L3 images. On L2 images, this makes sense and is well defined. On L3 images, it's less clear how to do this well, since the PSF is spatially variable and the input L3 image is likely a combination from different locations of input L2 images, and we don't want to track that level of detail.

Currently we make up a PSF for the L3 image by using a particular SCA. We should improve that in the following ways:

We should account for different pixel scales of the L3 image relative to the L2 image.
We should blur the L3 PSF by an additional box convolution according to the drizzle pixfrac
We should azimuthally average the PSF so that we are not including small-scale structure we don't really know anything about.

This should likely take the form of a new function in
https://github.com/spacetelescope/romancal/blob/main/romancal/lib/psf.py
that is analogous to create_gridded_psf_model, and likewise outputs a PSF model object that can be passed to fit_psf_to_image_model.