BUG: assert_series_equal wrong with check_category_order=False and NAs (GH#62008)

[jbrockmendel]

closes #62008

Missing values have code -1, and Index.take(-1) without allow_fill wraps around to the last category. assert_categorical_equal compared values via left.categories.take(left.codes), so two Categoricals holding equal values but storing their categories in a different order resolved the NA position to different labels and compared unequal:

values = ["B", np.nan, "D"]
left = pd.Series(pd.Categorical(values, categories=["B", "D"]))
right = pd.Series(pd.Categorical(values, categories=["D", "B"]))
pd.testing.assert_series_equal(left, right, check_category_order=False)
# AssertionError: Series category.values are different

This recodes both sides against the sorted categories and compares the codes, so -1 stays -1 and can only match -1.

Passing allow_fill=True, fill_value=np.nan to take — the approach in the closed GH-62011, GH-62017 and GH-66116 — fixes the reported case but casts integer categories to float64, so 2**53 and 2**53 + 1 collapse into one value and genuinely different Categoricals start comparing equal. There is a regression test pinning that.

Scope note: only assert_series_equal is affected on the public surface. assert_frame_equal and assert_index_equal do not accept or forward check_category_order, and the converse failure (unequal objects comparing equal) is not reachable through assert_series_equal, which compares values before it reaches assert_categorical_equal.

While here: check_category_order=False is still silently ignored when the categories cannot be sorted (mixed types), because the except TypeError fallback compares the categories in their original order. That is pre-existing and untouched here; the new tests pin the current behavior.

AI disclosure: drafted with Claude Code (claude opus 5 (high)), which wrote the fix and tests and ran the test suite; reviewed and revised by me.