fix: Handle inhomogeneous array shapes in to_arrays()

[dimitri-yatsenko]

Summary

• Fix to_arrays() failing when arrays have different but "almost compatible" shapes (e.g., (100,) vs (100, 1))
• Create empty object array and assign elements individually to avoid numpy broadcasting issues

Problem

When fetching blob attributes containing arrays of different shapes, to_arrays() would fail:

# Arrays with shapes (100,), (100, 1), (100, 2)
RandomNumbers.to_arrays('values1')  # ValueError

The fallback np.array(values, dtype=object) still fails because numpy tries to broadcast shapes that are "almost compatible".

Solution

Instead of np.array(values, dtype=object), create an empty object array and assign elements individually:

arr = np.empty(len(values), dtype=object)
for i, v in enumerate(values):
    arr[i] = v

This matches v0.14.7 behavior where fetch(*attrs) extracted columns from structured arrays with dtype=object, preserving individual array shapes.

Test plan

• Verified fix handles the reported case
• Add test case for inhomogeneous array shapes

Fixes #1380

🤖 Generated with Claude Code