Add universal image loader

[minhtien-trinh]

Features include:

• Dask array conversion for flexible chunking
• VRAM check to ensure memory sufficiency
• Zarr-backed optimization check for efficient handling of large images

This PR is part of the issue from spatialdata.

Note: This is a work in progress; initial testing is complete, but further validation is needed.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 35.87%. Comparing base (cff4896) to head (00d360e).
Report is 5 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #234      +/-   ##
==========================================
- Coverage   45.50%   35.87%   -9.64%     
==========================================
  Files          22       22              
  Lines        2149     2149              
==========================================
- Hits          978      771     -207     
- Misses       1171     1378     +207     

see 5 files with indirect coverage changes

[lucas-diedrich]

Hi! This PR and its related issue were referenced here.

We also work with large imaging datasets that exceed memory capacity, particularly CZI and whole-slide imaging files. A strategy that has worked well for us is:

1. Lazily loading data in small chunks using a custom image reader wrapped with dask.delayed,
2. Passing the chunks to dask.array (via dask.array.from_delayed), which is memory-mapped and avoids memory overflow,
3. Reassembling the chunks into a dask.array or Image2DModel.

You can find our implementation here.

For TIFF files, tifffile.memmap could potentially be used for lazy loading, as it returns a memory-mapped NumPy array. This approach does not extend to jpeg and png files. Would love to hear your thoughts!