Fix as_dataarray to apply coords parameter for DataArray input

[FBumann]

Fix as_dataarray to validate coords for DataArray input

Problem

When as_dataarray() received a DataArray with a coords parameter, the coords were silently ignored. This was inconsistent with all other input types (numpy, pandas, scalar) where coords are applied. Users passing DataArray bounds to add_variables with explicit coords would get no validation — mismatched coordinates would silently produce wrong results.

Solution

Add coordinate validation for DataArray inputs in as_dataarray():

• Shared dimensions must match exactly — raises ValueError on mismatch (no silent reindexing, which would introduce NaNs)
• Extra dimensions are rejected by default — raises ValueError if the DataArray has dims not in coords
• Missing dimensions are broadcast — dims in coords but not in the DataArray are added via expand_dims
• No coords means no validation — DataArrays pass through unchanged when coords=None

allow_extra_dims flag

Call sites that need xarray broadcasting for arithmetic opt in with allow_extra_dims=True:

Call site	allow_extra_dims	Why
add_variables (bounds, mask)	False (default)	Extra dims in bounds = shape mismatch bug
_multiply_by_constant	True	Broadcasting for arithmetic
Variable.to_linexpr	True	Broadcasting for coefficients
dot	True	Broadcasting for dot product

Performance

We introduce a new validation call on every as_dataarray call that passes a dataarray (ignored before this PR)
This should be fine performance wise imo and worth it.

Refactoring

• _coords_to_mapping(): Normalizes coords (sequence or mapping) to a dict. Replaces duplicated dict(zip(dims, coords)) in pandas_to_dataarray, numpy_to_dataarray, and the new DataArray branch.
• _validate_dataarray_coords(): Standalone validation function — normalizes coords, checks for extra dims, validates shared dim coords, expands missing dims. Single iteration over expected coords.

Incosistencies

• The coord validation is also done for pandas to dataarray, but there only a warning is raised instead of a ValueError
• broadcast mask currently has a FutureWarning and handles broadcasting differently. For missing dims it's consistent with as_dataarray. But it doesnt raise on misaligned coords. It fills with False. THis behaviour is kept (see 1bab496)

Examples

# Mismatched coords → ValueError
lower = xr.DataArray([0, 0, 0], dims=["time"], coords={"time": [0, 1, 2]})
m.add_variables(lower=lower, coords=[pd.RangeIndex(5, name="time")], name="x")
# ValueError: Coordinates for dimension 'time' do not match

# Extra dims → ValueError
da = xr.DataArray([[1, 2], [3, 4]], dims=["x", "y"])
as_dataarray(da, coords={"x": [0, 1]})
# ValueError: DataArray has extra dimensions not in coords: {'y'}

# Subset dims → broadcast
lower = xr.DataArray([0, 0], dims=["x"], coords={"x": [0, 1]})
m.add_variables(lower=lower, coords={"x": [0, 1], "y": [0, 1, 2]}, name="x")
# works, lower broadcast to shape (2, 3)

# No coords → pass through unchanged
as_dataarray(da)  # no validation, zero overhead

Checklist

• Code changes are sufficiently documented
• Unit tests pass (2077 passed)
• A note for the release notes
• I consent to the release of this PR's code under the MIT license

[FBumann]

Maybe remove explicit broadcast from masking added in a prior patch