TODO

o Implement ZarrADMatrixSeed and ZarrADMatrix objects (analog to
  H5ADMatrixSeed and H5ADMatrix objects from the HDF5Array package).

o Does Rarr provide an equivalent of h5ls() for quick inspection of
  a Zarr store? This would be very useful in the context of using
  things like ZarrSparseMatrixSeed() or ZarrSparseMatrix() where the
  user needs to identify the group nodes in a Zarr store that represent
  sparse matrices.

o See this PR for anndataR (Bioconductor package) by Mark Keller and
  Artür Manukyan: https://github.com/scverse/anndataR/pull/190

o In addition to looking at what Rarr has to offer to tackle the tasks
  below (array dimnames, groups, sparse arrays), take a look at
  pizzarr: https://github.com/keller-mark/pizzarr

o writeZarrArray() needs to write the array dimnames to disk.

o Add support for handling groups (or hierarchies of groups) of Zarr arrays.

o Improve support for Zarr v3. For example, one issue at the moment (Rarr
  1.11.24) is that Rarr::create_empty_zarr_array() only supports creation
  of Zarr v2 datasets. As a consequence, writeZarrArray(), which is based
  on Rarr::create_empty_zarr_array(), can only write in Zarr v2 format at
  the moment.
  - 2026/04/02 update: writeZarrArray() now has a 'zarr_version' argument
    that is set to 3 by default.

o Benchmarking:
  - Read/write speed compared with hdf5.
  - Impact of chunk geometry on read/write operations.
  - Concurrency: is it suported for reading and writing, or only for reading?
    (Note that hdf5 is notorious for not supporting concurrent writing to
    the same dataset.) Do some benchmarking using BiocParallel.


After ZarrArray is accepted and added to BioC 3.23
==================================================

o Implement saveZarrSummarizedExperiment() and
  loadZarrSummarizedExperiment() for saving/loading
  a Zarr-based SummarizedExperiment object to/from disk.