[ChrisW09]

Fixes #561

Problem

set_params(preprocessing_config=...) / set_params(preprocessing_config__...=...) prepare the next fit(). They replace estimator._preprocessor with a new, unfitted preprocessor. The fitted preprocessor that the in-memory model keeps using lives on estimator._data_module. Two code paths read estimator._preprocessor instead:

• build_save_bundle() saved the unfitted preprocessor, so the artifact's predict() raised NotFittedError after load();
• _prepare_continuation() built the continuation data with it, so fit(rebuild=False) raised the same error.

Fix

• deeptab/core/serialization.py (build_save_bundle): the artifact and its metadata take the data module's fitted preprocessor, falling back to estimator._preprocessor when there is none. preprocessor_kwargs and preprocessing_config are still saved as set, so a refit after load() uses the new settings. As a consequence, artifact_metadata["preprocessing"] holds the pending kwargs next to the fitted preprocessor's spec and fingerprint when settings changed after fit().
• deeptab/models/_mixins/fit.py (_prepare_continuation): continuation reuses the previous data module's fitted preprocessor, as the docs describe ("Continuation also reuses fitted preprocessing").

Tests

tests/integration/test_preprocessing_changes_after_fit.py (new), for MLPRegressor and MLPLSS:

• test_save_after_set_params_keeps_the_fitted_preprocessor: the reloaded artifact reproduces the in-memory predictions.
• test_continuation_after_set_params_uses_the_fitted_preprocessor

Both tests change cat_cutoff after fit(): on main, n_bins set through set_params() doesn't reach the new preprocessor (#489), while cat_cutoff does. All four cases fail on main and pass with this change.

Docs

No change needed.

Checks

• just test: 3240 passed, 46 skipped, 28 xfailed; coverage 91.28%
• just docs: build succeeded (warnings treated as errors)
• just check: all hooks passed (ruff, prettier, pyright)