DOC: Remove the spurious warnings from the Sphinx build

[bashtage]

The documentation build of main (121d869) ends with "build succeeded, 521
warnings" (533 in the last run on GitHub Actions, 37955470056; the footnotes of
the rank tests have been fixed since). With this change the same build ends
with "build succeeded, 3 warnings", with the packages of the CI run (Sphinx
9.1.0, sphinx-immaterial 0.13.9) on Linux with "-j 4", on Python 3.13 and 3.14.
The 3 that remain are in docstrings and belong to the docstring work: the type
list[str] in the docstrings of param_names in system/model.py.

Only doc/source/conf.py and doc/source/nitpick-exceptions change. The 521:

• NumPy scalars, 438 (numpy.float64 388, numpy.int64 48, numpy.int32 1,
  numpy.uint32 1). The array aliases expand in the signatures to
  ndarray[tuple[int, ...], dtype[float64]], and NumPy documents its scalar
  types as attributes, so the lookup of a class never succeeds. They are in
  nitpick-exceptions. The dtype stays in the signature, without a link.
• TypeAliasForwardRef 44 (188 in a serial build), pd.DataFrame 2 and pd.Series
  2. When a class docstring has a parameter without a type, napoleon in
  Sphinx 9 looks for the annotation of that name in the class and, as a side
  effect, merges napoleon_type_aliases into autodoc_type_aliases in place.
  From then on a nested annotation such as "DataFrame | None" is shown as
  "TypeAliasForwardRef('pandas.DataFrame') | None" with a warning that the
  class does not exist, and the dotted alias "pd.Series" turns pd into a
  pseudo module, so that "pd.DataFrame" cannot be evaluated and is shown as
  text. The signatures that are affected depend on the order in which the pages
  are built, which is why a parallel and a serial build disagree.
  napoleon_attr_annotations = False stops the lookup. The signatures are now
  shown like all others, and the types link to the pandas, NumPy and
  linearmodels documentation. The ndarray in the signature of LinearConstraint
  no longer links to IVData.ndarray.
• InvalidTestStatistic, 6. The alias pointed to linearmodels.shared, but the
  class is documented in linearmodels.shared.hypotheses.
• InapplicableTestStatistic, 4. It had no alias, and it is the type that
  pesaran_cd returns.
• DataFrame, 10. The signatures of the NamedTuples PanelModelData (8) and
  HausmanTestResults (2) show the bare name, because autodoc cannot evaluate
  the annotations of the new that NamedTuple generates. The
  python_type_aliases option of sphinx-immaterial expands it to
  pandas.DataFrame.
• csc_array, 1. Added with csr_array and coo_array to the aliases. The alias
  np.ndarray pointed to numpy.array and now points to numpy.ndarray.
• Private and undocumented classes, 8, which are in nitpick-exceptions: the
  _IVModelBase and _IVGMMBase in the annotations of the results classes (6),
  the protocol SupportsKeysAndGetItem of AttrDict.update (1), and the :class:
  role for functools.cached_property in changes/5.0.rst (1), which the Python
  documentation lists as a function. The exception for IVDataLike is removed
  because nothing refers to it any more.
• AttrDict, 3. get, popitem and setdefault are inherited from
  collections.abc.MutableMapping. Their docstrings are not numpydoc, and
  napoleon turns them into cross-references to types that do not exist. A
  handler of autodoc-skip-member leaves out the members of collections.abc, so
  these three are no longer documented as methods of AttrDict.
• list[str], 3, which stay. The docstring rules write "list of str".

napoleon_type_aliases and numpydoc_xref_aliases were copies of each other with
the same mistakes. They are now built from one dictionary. numpydoc is not in
extensions, so its options still have no effect. Aliases are added for
HausmanTestResults, AbsorbingLS and FamaMacBethResults, so that every estimator
and results class has one, and for callable, scalar and sequence, which the
docstring rules use and which resolve to the glossaries of Python and NumPy.

ipykernel 7 also prints "Kernel is running over TCP without encryption" once
for each notebook that nbsphinx runs, 12 in the CI log. They are not counted in
the 521, but they look like warnings. nbsphinx_execute_arguments sets the
kernel log level to ERROR.

To reproduce the CI build, install the packages of the CI run and run, from
doc/, "sphinx-build -M html source build -j 4". Pandoc is needed to convert
the notebooks.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.60%. Comparing base (121d869) to head (0a9c90b).
⚠️ Report is 10 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #767   +/-   ##
=======================================
  Coverage   99.60%   99.60%           
=======================================
  Files         127      127           
  Lines       23832    23832           
  Branches     1783     1783           
=======================================
  Hits        23737    23737           
  Misses         35       35           
  Partials       60       60           

Flag	Coverage Δ
adder	99.58% <ø> (ø)
subtractor	99.58% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.