Improve the shuffling/preprocessing code

The current shuffling code is very slow, reportedly. The first step should be profiling the code especially in the wake of https://github.com/scverse/anndata/pull/2005. Other things should be considered in light of profiling

• Try out https://zarrs-python.readthedocs.io/en/stable/ for the writing if that appears to be the bottleneck
• Trying out a non-dask implementation instead of dask for loading X so we don't have to deal with whatever overhead is there
• Understand if tiny shards obviate the need for a complete reshuffle, in which case https://github.com/laminlabs/arrayloaders/blob/main/arrayloaders/io/store_creation.py#L16-L51 will likely suffice and we can simply re-write the data with tiny shards

Start by profiling the shuffling and preprocessing code, especially in light of scverse/anndata pull request 2005. Compare the listed options, including zarrs-python for writing, a non-dask X loader, and the tiny-shard approach in arrayloaders/io/store_creation.py. Done means the bottleneck is identified and a measured direction is established.