Add Zarr v3 output and per-dimension sharding to rechunk2zarr

[Copilot]

rechunk2zarr only produced Zarr v2 stores. This adds a --zarr-format switch and per-dimension shard sizes so users can emit v3 sharded arrays.

Changes

• rechunk.py
  • New rechunk_zarr kwargs: zarr_format (2 or 3, default 2) and t_shard / c_shard / spatial_shard.
  • New _validate_shards helper: rejects shards when zarr_format != 3, requires each shard to be a positive-integer multiple of its chunk, and requires an explicit c_chunk when c_shard is set.
  • Passes shards= to create_array; axes without an explicit shard collapse to one chunk per shard (i.e. no sharding on that axis).
  • Single-pass write loop now iterates by shard size when sharding is enabled so each shard is written exactly once (buffer grows with shard, not chunk, size).
  • When writing to a v3 destination, compressors="auto" is used — v2 numcodecs compressors from the source aren't valid v3 BytesBytesCodecs and can't be passed through.
• CLI: --zarr-format {2,3}, --t-shard, --c-shard, --spatial-shard.
• test_rechunk.py (new): validation edge cases + end-to-end tests for v2, v3-no-shards, v3 full sharding, and v3 partial sharding.
• README.md: documents the new flags and library usage.

Example

from rechunk import rechunk_zarr

rechunk_zarr(
    src_path="myfile.zarr",
    dst_path="myfile_v3.zarr",
    t_chunk=24, c_chunk=8, spatial_chunk=100,
    zarr_format=3,
    t_shard=48,          # multiple of t_chunk
    spatial_shard=500,   # multiple of spatial_chunk; c axis stays unsharded
)

rechunk2zarr src.zarr dst.zarr --zarr-format 3 \
    --t-chunk 24 --t-shard 48 --spatial-chunk 100 --spatial-shard 500

• Fixes Rechunk into Zarr v3 #34

[tackandr]

@copilot When I open a resulting dataset it says: KeyError: 'Zarr object is missing the dimension_names metadata which is required for xarray to determine variable dimensions.'

[Copilot]

@copilot When I open a resulting dataset it says: KeyError: 'Zarr object is missing the dimension_names metadata which is required for xarray to determine variable dimensions.'

Fixed in f03d3b4. The v3 destination arrays weren't getting a dimension_names metadata entry — v2 stores dims in the _ARRAY_DIMENSIONS attribute, which xarray doesn't consult for v3 stores. rechunk_zarr now reads the source's dimension names (from v3 metadata or the v2 _ARRAY_DIMENSIONS attr) and passes them to create_array(..., dimension_names=...) for every v3 array (data + coords), so xr.open_zarr(dst) works.