[ChrisW09]

Fixes #557

Problem

BaseDistribution.__init__ stored the "none", "square", "probabilities" and "log" parameter transforms as lambdas, and get_transform() fell back to another lambda. Lambdas defined inside a function can't be pickled, so pickling failed with Can't get local object 'BaseDistribution.__init__.<locals>.<lambda>' for:

• every distribution family;
• every fitted LSS estimator;
• anything that pickles the LSS training module, such as Lightning's ddp_spawn strategy.

Fix

deeptab/distributions/base.py replaces the lambdas with module-level functions (_identity, _square, _softmax_last_dim, _stable_log) that compute exactly the same values. The get_transform() fallback uses _identity. The behaviour is unchanged otherwise. Unknown transform names still map to the identity; that's a separate question.

Tests

In tests/integration/test_save_load.py, the module that already holds the pickle tests and has the S301 ruff exemption:

• test_every_distribution_family_survives_a_pickle_round_trip[<all 16 registered families>]: same class and parameter names after pickle.loads(pickle.dumps(...)). For normal, the restored family also produces the same transformed parameters.
• test_fitted_lss_can_be_pickled[normal|quantile|categorical]: a fitted MLPLSS pickles and unpickles as an unfitted estimator, as documented for the other estimators.

All 19 fail on main and pass with this change. I confirmed the ddp_spawn failure on main (MLPLSS(...).fit(..., accelerator="cpu", devices=2, strategy="ddp_spawn") raises the same AttributeError). In my sandbox, an end-to-end ddp_spawn run can't complete for any estimator, including MLPRegressor, because torch_shm_manager can't bind. With this change the LSS run gets past pickling to that same environment error.

Docs

No change needed.

Checks

• just test: 3255 passed, 46 skipped, 28 xfailed; coverage 91.28%
• just docs: build succeeded (warnings treated as errors)
• just check: all hooks passed (ruff, prettier, pyright)