Confirm Filter Downsample Scaling Behavior

[garrettwrong]

When going through the filter code converting to stacks I noticed that the code may be adjusting for downsampling in two ways and Im not sure it should do both.

One scaling is via the pixel_size, which scales bandwidth applied to radial grid inside phase flipping. The radial grid is derived from the omega passed from evaluate.

Another scaling occurs via filter.scale which scales omega directly (before passing to filter.evaluate).

We generally get best results with phase flipping at full image size. I think it is worth reviewing the code for correctness in case this is why the phase flipping after downsample under performs expectations.

[garrettwrong]

This is not correct:

# pixel_size has already been adjusted in the ImageSource and passed
# to `im`, so we instantiate the new Image with im.pixel_size.
    return Image(im_ds, pixel_size=im.pixel_size).stack_reshape(

Scaling pixel_size correctly here in combination with removing the double scaling from the original issue resolved the bugs I found (both the pf/ds vs ds/pf and caching issues) but breaks the situation that hack was originally put in place for. So it looks like in hacking around that pixel_size bug we fixed one minor thing and broke two major things.

[garrettwrong]

Continuing working on this. The attempted fixes in #1302 were really over-working things in multiple files/components. I believe I've found the line causing the original bug for #1318 using some basic detective work 👨‍🔬 . Cleaning up. Should have it together before I head out this week so it can be merged after the other filter work and get tested on the recons etc for the release.

[garrettwrong]

(Adding missing context discussed in our dev meeting). Fixing the bug reported here in 1395 by removing the redundant scaling and adding additional unit tests to verify combinations of downsample phaseflip and cache indicated there were still underlying bugs relating to incorrect pixel_size.