on-the-fly PSFs have a subpixel offset relative to precomputed library PSFs

[kvangorkom]

When the Pandeia library PSFs are generated, the requested offsets for the PSFs are forced to nearest integer upsampled pixel, and then recentered by a simple pixel remapping after the call to WebbPSF (so the convolution kernel created from the PSF is centered). For on-the-fly PSFs, I've intentionally avoided forcing the offset to an integer multiple of the upsampled pixel, since we care (particularly for MIRI) about small changes to these spatially-dependent PSFs. However, the pixel remapping approach then has the side effect of producing a slightly off-center convolution kernel.

Possible mitigation strategies: interpolate the PSF for the centering step (I really don't want to do this), or adopt a higher upsampling value.