NaN in user-supplied data is silently filled instead of surfaced

[FBumann]

Issue Description

When user-supplied data contains NaN — usually a sign of missing or bad input — linopy silently fills it with an operator-specific neutral element rather than surfacing it. The fill value differs per operator, and in every case a data error that should stop the run is absorbed without a trace.

Reproducible Example

import numpy as np
import pandas as pd
import xarray as xr
import linopy

m = linopy.Model()
x = m.add_variables(coords=[pd.RangeIndex(3, name="time")])
data = xr.DataArray([1.0, np.nan, 3.0], dims=["time"], coords={"time": range(3)})

print((x + data).const.sel(time=1).item())             # 0.0 — NaN -> 0
print((x * data).coeffs.squeeze().sel(time=1).item())  # 0.0 — NaN -> 0 (variable zeroed)
print((x / data).coeffs.squeeze().sel(time=1).item())  # 1.0 — NaN -> 1 (variable unchanged)

The NaN at time=1 is replaced — 0 for +, 0 for * (zeroing the variable out), 1 for / (leaving it unchanged). No exception, no warning.

Expected Behavior

NaN in user-supplied data should not be silently substituted. The user should say what it means (data.fillna(0), factor.fillna(1), …) — or linopy should raise.

Confirmed on linopy 0.7.0. Part of the arithmetic-convention work (#591).