Fix sharded arrays losing their shard config on axis changes

[TomNicholas]

Closes #1076.

A chunk manifest entry locates a whole shard, but several operations in manifests/ read ArrayV3Metadata.chunks for that shape — which zarr defines as the inner chunk shape when a sharding codec is present (the outer chunk grid is .shards). The parsers and the icechunk writer already get this right; the array-op layer did not.

The visible symptom was that concatenating sharded virtual arrays along a new dimension raised:

ValueError: The shard's `chunk_shape` and array's `shape` need to have the same number of dimensions.

because stack grew the array's ndim without touching the shard config. But the same conflation had quieter consequences: stack/broadcast_to wrote an inner chunk shape back as the outer chunk grid, and a no-op slice of a sharded array failed to short-circuit because it compared a chunk-grid selector derived from inner chunks against the shard-shaped manifest grid.

What's in here

• manifest_chunk_shape(metadata) — one accessor for "the region one manifest entry locates", used in stack, broadcast_to, concatenate's partial-chunk check, _missing_element_mask, apply_selection, the reindex path, check_combinable_zarr_arrays, ManifestArray.empty_like, and the oversized-chunk warning.
• copy_and_replace_metadata realigns any sharding_indexed codec's inner chunk_shape when the outer chunk shape gains or loses length-1 axes. Both directions matter: stack/expand_dims/broadcast_to add an axis, integer indexing drops one (only legal where the chunk length is already 1). Nested shards recurse.

Realigning length-1 axes is byte-safe — it leaves the inner chunk grid, the C-order shard index layout, and every inner chunk's bytes untouched — so existing references stay valid rather than needing a rewrite.

Tests

Written first, all red before the fix:

• TestCopyAndReplaceMetadataSharding — prepended/middle/multiple added axes, dropped axis, nested shards, the ambiguous-alignment case where old_chunks already contains 1s, unsharded codecs left alone, and rejection of a chunk shape that isn't an axis change.
• TestSharded in test_manifests/test_array.py — expand_dims, stack, broadcast_to, concatenate, plus the indexing paths (no-op full slice, integer index dropping an axis, shard-aligned slice).
• test_concat_sharded_arrays_along_new_dim_roundtrip_icechunk — the end-to-end payoff: two sharded Zarr stores, virtualized, concatenated along a new dim, written to Icechunk, and read back element-for-element equal to xr.concat of the originals. That's what makes the byte-safety claim concrete rather than asserted.

Full suite: 642 passed, no new mypy errors.

Why it came up

Building an Icechunk datacube from an archive of per-forecast-hour .zarr.zip files (via the new ZippedZarrParser), where every variable is a single (720, 1440) shard of (45, 45) chunks — so every variable tripped this the moment the files were concatenated along init_time/forecast_hour.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.64%. Comparing base (1ba8470) to head (e8c947c).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1077      +/-   ##
==========================================
+ Coverage   89.60%   89.64%   +0.03%     
==========================================
  Files          41       41              
  Lines        2664     2674      +10     
==========================================
+ Hits         2387     2397      +10     
  Misses        277      277              

Files with missing lines	Coverage Δ
virtualizarr/manifests/array.py	86.79% <ø> (ø)
virtualizarr/manifests/array_api.py	98.59% <100.00%> (ø)
virtualizarr/manifests/indexing.py	92.42% <100.00%> (ø)
virtualizarr/manifests/store.py	90.38% <100.00%> (ø)
virtualizarr/manifests/utils.py	93.24% <100.00%> (+1.05%)	⬆️

🚀 New features to boost your workflow:

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