Keep lazy_elementwise blocks as arrays for scalar-returning ops (#6965)

[gaoflow]

🚀 Pull Request

Description

Closes #6965.

Computing (e.g. saving) a scalar lazy cube whose units had been converted failed:

import iris.analysis, dask, numpy as np, dask.array as da
from iris.coords import DimCoord
from iris.cube import Cube

cube = Cube(da.arange(10), units="m",
            dim_coords_and_dims=[(DimCoord(np.arange(10), var_name="x"), 0)])
cube = cube.collapsed("x", iris.analysis.MIN)   # -> 0-dimensional, lazy
cube.convert_units("km")

delayed = iris.save(cube, "test.nc", compute=False)
dask.compute(delayed)

AttributeError: 'float' object has no attribute 'size'

The error only appears with convert_units and only when the result is stored lazily (calling cube.data directly is fine).

Cause

convert_units applies the conversion lazily via iris._lazy_data.lazy_elementwise, which maps the operation over the blocks of the lazy array. The operation is cf_units.Unit.convert, and for a 0-dimensional block that call returns a Python float rather than a 0-d array. Dask's store step then does block.size, which a float does not have.

Fix

Wrap each block result in np.asanyarray inside lazy_elementwise, so a block always remains an array regardless of what the operation returns. For a normal (n-d) block this is a no-op; for the scalar case it restores the expected 0-d array. Masks are preserved (asanyarray).

Verification

• New test Test_lazy_elementwise::test_scalar_returning_op_on_0d — an op that returns a Python scalar for a 0-d block now yields a 0-d ndarray. Fails on main, passes here.
• Reproducer above now saves successfully and round-trips the correct value (0.0 km); a 1-d convert_units remains lazy with correct values.
• Full tests/unit/lazy_data/ suite (51) and the convert_units cube tests pass.
• ruff check / ruff format clean.