Only use oindex for advanced indexing (int arrays)

This change means that more of the array API is supported for
storage backends that don't have oindex (e.g. the new Zarr v3
implementation).

cubed/core/ops.py

@@ -489,6 +489,7 @@ def merged_chunk_len_for_indexer(s):
         extra_projected_mem=extra_projected_mem,
         target_chunks=target_chunks,
         selection=selection,
+        advanced_indexing=len(where_list) > 0,
     )
 
     # merge chunks for any dims with step > 1 so they are
@@ -509,10 +510,19 @@ def merged_chunk_len_for_indexer(s):
     return out
 
 
-def _read_index_chunk(x, *arrays, target_chunks=None, selection=None, block_id=None):
-    array = arrays[0]
+def _read_index_chunk(
+    x,
+    *arrays,
+    target_chunks=None,
+    selection=None,
+    advanced_indexing=None,
+    block_id=None,
+):
+    array = arrays[0].zarray
+    if advanced_indexing:
+        array = array.oindex
     idx = block_id
-    out = array.zarray.oindex[_target_chunk_selection(target_chunks, idx, selection)]
+    out = array[_target_chunk_selection(target_chunks, idx, selection)]
     out = numpy_array_to_backend_array(out)
     return out