test: isolate test_pybamm_import from sys.modules mutation (fixes #5402)

[gaoflow]

Isolate test_pybamm_import and test_import_optional_dependency from sys.modules mutation

Fixes #5402.

Problem

tests/unit/test_util.py::TestUtil::test_pybamm_import mutates sys.modules
in-process to pretend that all optional dependencies are absent, then unloads
pybamm itself so the next importlib.import_module("pybamm") re-runs the
top-level import logic. The hand-rolled try/finally restores entries
afterwards, but for the duration of the test the pytest worker is in a
partially-consistent state, and any partial import state cached during the
re-import (logging handlers, import locks, optional-dependency fallbacks
already evaluated in other modules) cannot be undone by writing back to
sys.modules.

With pytest-xdist that surfaces as order-dependent CI flakiness in
unrelated tests — the issue author confirmed in #5402 that simply
@pytest.mark.skip-ing this single test made all CI tests pass consistently.
@pytest.mark.forked did not help because fork() inherits the parent's
already-cached import state.

Fix

Remove the in-process mutation entirely.

• test_pybamm_import now runs the import pybamm check inside a fresh
  subprocess.run([sys.executable, "-c", ...]). The script masks every present
  optional dependency (and any already-imported sub-modules) inside its own
  throw-away interpreter, then imports pybamm. The parent worker's
  sys.modules is never touched. A non-zero exit code propagates stdout
  and stderr into the pytest failure message so debugging stays easy.

  This is strictly stronger isolation than pytest.mark.forked: a spawned
  Python interpreter starts with a clean slate, while fork() copies the
  parent's already-cached import state.

• test_import_optional_dependency uses pytest's monkeypatch.setitem
  to mask sys.modules entries. The mutation is automatically reverted by
  pytest's fixture teardown even on failure — there is no need for this
  test to run in a subprocess because it does not unload pybamm.

Verification

The refactored tests preserve the exact original intent:

• test_pybamm_import still asserts that import pybamm succeeds when every
  present optional dependency is marked missing.
• test_import_optional_dependency still asserts that
  pybamm.util.import_optional_dependency(pkg) raises
  ModuleNotFoundError with the expected message.

$ pytest tests/unit/test_util.py -v
...
tests/unit/test_util.py::TestUtil::test_pybamm_import PASSED
tests/unit/test_util.py::TestUtil::test_import_optional_dependency PASSED
tests/unit/test_util.py::TestUtil::test_optional_dependencies PASSED
tests/unit/test_util.py::TestUtil::test_fuzzy_dict PASSED
tests/unit/test_util.py::TestUtil::test_get_parameters_filepath PASSED
tests/unit/test_util.py::TestUtil::test_is_constant_and_can_evaluate PASSED
tests/unit/test_util.py::TestSearch::test_url_gets_to_stdout PASSED
============================== 7 passed in 4.15s ===============================

Negative-test sanity (deliberately mask numpy instead of an optional dep):

$ python -c "import subprocess, sys, textwrap; ..."
returncode: 1
stderr tail: ['ModuleNotFoundError: import of numpy halted; None in sys.modules']

so the subprocess approach still surfaces a regression if pybamm ever
starts to require an optional dep at import time.

ruff check tests/unit/test_util.py and ruff format --check tests/unit/test_util.py are both clean.

Why this scope

The smallest change that fixes the root cause: only test_util.py is touched.
No production code is altered. The two adjacent tests that exhibit the same
in-process-mutation anti-pattern are fixed together so the file is internally
consistent.