Add multidimensional drift correction and strip-based alignment - EDS, 4DSTEM prior to arXiv

[bobleesj]

What problem this PR addreseses

Support 3D, 4DSTEM drift correction.

PRs broken down as shown below:

• #14 Split Drift into a package
• #15 Add affine drift diagnostics
• #17 Correct spectrum-image drift
• #18 Correct 4D-STEM scan axes
• #20 Accelerate affine drift search
• #21 Reproduce manuscript workflows
• #22 Simplify the DriftCorrection API
• #23 Clarify the preprocessing module

Before

src/quantem/imaging/
├── drift.py              1,832 lines
└── drift_utils.py          664 lines

After

├── correction.py         orchestration
├── preprocess.py        input preparation
├── apply.py              dataset propagation
├── diagnostics.py        numerical diagnostics
├── report.py             structured summary
├── io.py                 EMD metadata and pairing
├── fourdstem.py           4D-STEM collection logic
├── plot.py               visual diagnostics
└── core/
    ├── affine.py
    ├── strip.py
    ├── nonrigid.py
    ├── knots.py
    └── warping.py

Public API change

existing usage	proposed usage	why
from_data(...)	from_images(...)	2D images - different from from 4dstem, 3d dataset
align_translation()	Same	translation is align
align_affine()	correct_affine()	affine is correcting knots
align_nonrigid()	correct_nonrigid()	non-rigid is correcting knots
generate_corrected()	corrected()	concise, pythonic
number_knots, step	num_knots, max_drift_rate	clarifies the model count and physical search meaning

Affine speed:

metric	scipy full, cpu (2016 paper)	torch full, cuda	torch automatic pyramid, cuda
algorithm time	2439.956 s	5.788 s	0.441 s
speedup vs scipy	1×	421.5×	5536.0×
drift rate, row (px/line)	-0.073000	-0.073000	-0.072461
drift rate, col (px/line)	0.139000	0.139000	0.139648

The pyramid downsampling strategy enables real-time drift correction on any personal laptops.

What should the reviewer(s) do

Tutorials are available here: https://github.com/bobleesj/quantem-tutorials/tree/drift-tutorial/tutorials/imaging/drift
.

[bobleesj]

@wwmills ready for review.

If you agree with the API proposed and folder architecture, we can probably merge so that you can add your new algo as needed and refactor with upcoming PRs. This PR is to enable further scaling of features and reproduce manuscript results.