fix: support mixed slice indexing in ChunkCachedArray

[mvanhorn]

Description

Fixes ChunkCachedArray._vindex_get for mixed vectorized keys (integer arrays plus slices), which xarray produces when some dimensions are indexed and others are left whole.

A Delft3D surface field loaded with to_chunk_cached_arrays() failed in _gather_corners: the array is (time, mockZ, N, M) with a singleton mockZ, time/N/M are indexed, and mockZ is left as slice(None). _raw_vindex treated that slice as a 1-D integer array during negative-index normalization. The same simulation works with windowed arrays, so this is local to the cache backend.

Changes, all in _vindex_get:

• Broadcast the integer index arrays to a common shape and expand each slice with slice.indices for its dimension.
• Order the slice axes after the broadcast index axes, matching xarray's vectorized-indexing order. Then flatten into per-dimension coordinates for the existing chunk-grouping _raw_vindex, and reshape the result.
• Keep the existing fast path for equal-length 1-D all-array keys.
• Return early for zero-point selections, and raise IndexError for integer indices out of range before chunk lookup. Valid negative indices still work.

The per-chunk LRU fetch path is unchanged. The bug reproduces with synthetic arrays, so the tests don't need the reporter's data.

Closes #2897

Checklist

• Closes: TypeError in ChunkCachedArray: '<' not supported between instances of 'slice' and 'int' #2897
• Tests added
• This PR targets the correct branch (main for normal development, v3-support for v3 support)

AI Disclosure

• This PR contains AI-generated content.
  • I have tested any AI-generated content in my PR.
  • I take responsibility for any AI-generated content in my PR.

I used AI assistance to write the fix and tests. tests/test_chunk_cached_array.py and tests/test_interpolation.py pass locally (59 passed, Python 3.13).

[VeckoTheGecko]

Hi @mvanhorn . Can you please complete the AI disclosure part of the PR template? We want to make sure all PRs submitted to our project have a human behind the wheel.

[mvanhorn]

Done, the AI disclosure section is filled in. I used AI assistance for the fix and tests, reviewed them, and ran test_chunk_cached_array.py and test_interpolation.py locally (59 passed). I also dropped a uv.lock that had slipped into the branch.

[VeckoTheGecko]

superceded by #2921