Use shared backend test structure in imputation tests, fix DropMissingData output type

[solegalli]

Migrates the imputation tests to the shared test structure introduced in #1045 (merged), and fixes a DropMissingData bug that the new output-type checks caught.

Tests

• tests/test_imputation/conftest.py: data_na (shared by five test files) and data_na_dob (the same data plus a never-null datetime column, for DropMissingData and MissingIndicator), as fixtures returning plain dicts.
• File-local helpers (_cols, _null_count, _values, _same_values, assert_df_equal, _missing_count, _to_list, _make_series) replaced by the shared make_df fixture, to_dict, null_count and make_series.
• Every transform output is now checked to be of the input backend with isinstance(X, make_df).
• test_arbitrary_imputer.py now uses the shared data, which adds the Studies column, so its n_features_in_ assertions are 5.
• RandomSampleImputer's three seed-per-observation tests are merged into one parametrized test.

Bug fix: DropMissingData
_select_rows() receives the narwhals frame returned by check_X, but dispatched on nwd.is_pandas_dataframe(X), which is never true for a narwhals frame (narwhals warns about it). As a result:

• when no variable was selected (e.g. missing_only=True on a clean training set), transform() returned the narwhals frame instead of a pandas/polars dataframe;
• the benchmarked pandas fast path never ran, so pandas input silently used the slower narwhals path.

It now branches on X.implementation.is_pandas() and always returns the native frame.

RandomSampleImputer: per-observation seed (#601)
With seed="observation", the seed of each observation is now a hash of the values of the variables in random_state, instead of their rounded sum or product. As a result:

• observations with the same values in those variables receive the same imputation, regardless of their position in the dataframe (also with duplicated pandas index labels);
• negative or very large values no longer raise an error, and different observations no longer share a seed because of rounding;
• missing values in the seeding variables count as 0;
• seeds are computed from the data before any variable is imputed, the same way for pandas and polars.

The seeding_method parameter is removed: with hashing there is nothing to add or multiply. Imputed values with seed="observation" differ from previous releases. The user guide is updated accordingly.

Closes #601

Verification

• tests/test_imputation: 237 passed; the 7 failures are the pre-existing sklearn estimator checks, identical to narwhals-migration.
• Full tests/ suite: same failure set as narwhals-migration, nothing new. flake8 and mypy clean.