Fix 3D to 2D phase reconstruction

[talonchandler]

Fixes ##172. Moved from old recOrder draft PR.

This PR fixes the 3D to 2D phase reconstruction, best demonstrated in this video with OPS006 data:

Screen.Recording.2025-05-07.at.5.38.07.PM.mov

The key fixes were:

• rewriting the reconstruction with an SVD-based method, which synchronizes this code path with the rest of the waveorder framework. This allowed me to debug with easier math.
• turning off the default background filter, which was too aggressive and accidentally suppressed the signals.

While debugging I made two observations:

• the existing code only supported CLI reconstructions with data where the in-focus slice was in the central slice. For the OPS data, a single-slice offset of focus (25 um!) was enough to ruin the 3D-to-2D reconstruction. I added a new z_focus_offset parameter.
• if the NA_illumination = NA_detection, the 3D-to-2D phase transfer function goes to zero. I need to think through it this makes sense, or if this is indicating a deeper modeling issue.

TODO:

• save absorption to the output zarr, with correct channel names. I'm estimating absorption, just not saving it for now skipped for now...will reevaluate after testing with absorbing sample.
• consider precalculating the SVD with the transfer function. The SVDs are reasonably large (a 2 x Z matrix needs an SVD for each X*Y pixel), so I expect this to be the bottleneck. DONE.

Paths:

• /hpc/projects/comp.micro/microscope_dev/2025-3d-to-2d-debug/2025-05-07-3d-to-2d-debug/visualize.py will open napari like in the viewer above

[ziw-liu]

For the OPS data, a single-slice offset of focus (25 um!) was enough to ruin the 3D-to-2D reconstruction.

Does this mean that this method will be very sensitive to tilted sample? @ahillsley can confirm but I recall seeing some variation of focus within FOVs. Maybe they will need tiling here?

[talonchandler]

Maybe they will need tiling here?

Very possible. On our first few iterations we should definitely keep an eye on our reconstruction quality that varies with focus across the FOV.

[ziw-liu]

Can this be moved to tests? We don't have doctest set up.

[ziw-liu]

Great work!

[ziw-liu]

This works much better with cellpose!