fix: serialize MultiIndex level names as JSON for scipy netCDF backend

[FBumann]

Summary

• Fixes ci: FAILED test/test_io.py::test_model_to_netcdf_with_multiindex - KeyError: ('U', 24) #525: to_netcdf fails with KeyError: ('U', 24) when only scipy is installed (no netCDF4 / h5netcdf). xarray falls back to scipy's netCDF3 backend, which cannot write unicode-array (<U_) attributes — the format used by linopy to store MultiIndex level names as a Python list of strings.
• Encode the level names as a JSON-encoded scalar string instead, mirroring the existing pattern in the same file for _relaxed_registry / _piecewise_formulations. read_netcdf still accepts the legacy list form for backward compatibility with files written by older linopy versions.
• Drops a now-dead skipif on the existing MultiIndex test (the xarray bug it guarded was fixed in xarray==2024.02.0, and pyproject.toml already pins xarray>=2024.2.0).

Repro (without this patch, on any xarray version including current 2026.4.0)

import pandas as pd
from linopy import LESS_EQUAL, Model

m = Model()
idx = pd.MultiIndex.from_tuples(
    [(1, "a"), (1, "b"), (2, "a"), (2, "b")], names=["first", "second"]
)
x = m.add_variables(4, pd.Series([8, 10, 12, 14], index=idx), name="x-var")
y = m.add_variables(0, pd.DataFrame([[1, 2], [3, 4], [5, 6], [7, 8]], index=idx), name="y-var")
m.add_constraints(x + y, LESS_EQUAL, 10, name="c1")
m.add_objective(2 * x + 3 * y)
m.to_netcdf("test.nc", engine="scipy")  # KeyError: ('U', 24)

The bug is a scipy-backend limitation, not an xarray regression — upgrading xarray does not fix it.

Test plan

• test_model_to_netcdf_with_multiindex_scipy_engine — exercises engine="scipy" explicitly and asserts the on-disk *_multiindex attribute is a scalar string (locks in the actual fix, not just the round-trip).
• test_read_netcdf_with_multiindex_legacy_list_attr — backward-compat: rewrites a netCDF file with the old list-of-strings attribute and confirms read_netcdf still parses it. Skipped when netCDF4 isn't installed (scipy cannot produce that format).
• Existing test_model_to_netcdf_with_multiindex still passes (default engine).
• pytest test/test_io.py — 25/25 pass.
• ruff check / ruff format --check / pre-commit clean.

Notes

Debatable if we need this i think