fix: apply drop_axes squeeze in partial decode path for sharding (#3691)

[abishop1990]

Summary

Fixes #3691.

Mixed integer/list indexing on sharded arrays (e.g. arr[0:10, 0, [0, 1]]) raised:

ValueError: could not broadcast input array from shape (10,1,2) into shape (10,2)

Root Cause

When OrthogonalIndexer processes advanced indexing (slices + integers + arrays), it applies ix_() to chunk_selection to set up orthogonal numpy indexing. Integer indices become 1-element ranges (size-1 dimensions) via ix_().

CodecPipeline.read_batch() has two paths:

1. Non-partial decode (regular codecs): Applies drop_axes.squeeze() to remove size-1 integer dims ✅
2. Partial decode (ShardingCodec): Missing drop_axes.squeeze() ❌

ShardingCodec._decode_partial_single() receives the ix_()-transformed chunk_selection, which looks like pure fancy indexing to get_indexer(), so it routes to CoordinateIndexer. The result is reshaped to the broadcast coordinate shape (10, 1, 2) instead of (10, 2).

Fix

Apply drop_axes squeeze to chunk_array in the partial decode branch of read_batch(), matching the non-partial path behaviour:

if drop_axes != ():
    chunk_array = chunk_array.squeeze(axis=drop_axes)
out[out_selection] = chunk_array

Testing

Added test_sharding_mixed_integer_list_indexing that verifies:

• Shape and data equality between chunked and sharded arrays for mixed indexing
• Multiple integer axes (arr[0, 0, [0, 1, 2]])
• Slice + integer + slice (arr[0:5, 1, 0:3])

tests/test_codecs/test_sharding.py  125 passed, 1 skipped
tests/test_indexing.py  149 passed, 1 skipped, 5 xfailed

[d-v-b]

are we sure that drop_axes is always a tuple? maybe it is, but if not, might be safer to check if its length is 0

[abishop1990]

Good catch — switched to if drop_axes: (truthy check) throughout codec_pipeline.py. It's both more Pythonic and type-agnostic, even though drop_axes is annotated as tuple[int, ...] everywhere. Pushed in the latest commit.