perf: dont use a vstack before indexing

[ilan-gold]

@felix0097 Here is the no vstack. I don't want to merge it yet because it is generic over sparse and dense and with sparse, it doesn't help (and is more complicated). An overview of the two pipelines for comparison (after data fetching):

main

What's on main now will stack together the fetched chunks and then yield from the vstacked result batch_size subsets. If preload_gpu is enabled, the vstacking occurs just after the data is loaded onto the GPU

this branch

With this branch, the chunks are either left alone or converted to the GPU if preloading is enabled. Then they are yielded from based on batch_size.

I put this on a branch first to make sure it benefits dense. If it does, I'll put the feature behind a flag, and we can turn it on for dense only (or sparse if I have missed something perf-wise)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 94.91%. Comparing base (2ef3b8a) to head (623bbe7).

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #25      +/-   ##
==========================================
+ Coverage   94.71%   94.91%   +0.19%     
==========================================
  Files           7        7              
  Lines         511      531      +20     
==========================================
+ Hits          484      504      +20     
  Misses         27       27              

Files with missing lines	Coverage Δ
arrayloaders/io/zarr_loader.py	95.01% <100.00%> (+0.35%)	⬆️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

Deployment URL: https://7fc446fa.arrayloaders.pages.dev