fix: state-conditional propensity head + posterior-mixing correctness (v0.1.0)

[gmgeorg]

Squashes the fix/1 branch's 8 commits into one for review. Summary of what changed and why:

Core correctness fix

• OutcomeLoss mixed predictive states using the prior weights P(s_k|x) instead of the posterior P(s_k|x,a); since treatment a is observed, only the posterior makes TreatmentLoss + OutcomeLoss telescope exactly into -log p(a,y|x). Computing the posterior requires each state's own P(a|s_k,x), so the propensity head is now state-conditional (one BiasOnly+sigmoid per state, like the outcome head) instead of pre-mixed into a single marginal column -- a breaking change to the model's output shape; models trained/saved with <0.1.0 are not compatible with this release.
• log(weights + eps) in the state-mixture losses clamped log-weights at log(eps), distorting sharp/near-degenerate predictive states; replaced with a safe_log that only substitutes for exact-zero entries.
• CausalLoss's alpha (treatment-loss weight) was a plain Python float. model.fit() traces CausalLoss.call() into a compiled graph (run_eagerly=False by default), so the float got baked in as a constant at the first trace; AlphaScheduleCallback rebinding loss._alpha between epochs only updated the eager Python attribute, which the compiled training function never read again -- the annealing schedule silently never reached the optimizer, despite reporting the intended curve. alpha is now a tf.Variable, updated via .assign(...). Verified directly (tf.function trace before/after) and end-to-end via model.fit().
• predict_ute_binary/predict_ute_continuous hardcoded n_outcome_pred_cols=2 (Normal-only), silently mis-slicing columns -- and computing the wrong number of states -- for any other outcome distribution (e.g. the exponential/survival model), returning a plausible-looking but meaningless number with no error. Now converts each outcome distribution's raw parameters to its actual mean (Normal: loc; Exponential: exp(-log_rate)), raising NotImplementedError for any distribution it doesn't know how to convert.
• build_toy_model/build_model_binary_normal didn't compile causal_loss_metric (only build_model_binary_exponential did), so get_default_callbacks' documented recommended monitor ("val_causal_loss_metric") didn't exist for them; Keras only warns, never raises, on a missing monitored metric, so EarlyStopping/ReduceLROnPlateau silently never triggered.

New helpers

• utils.get_column_layout(model) / utils.ColumnLayout: reads n_outcome_pred_cols/n_treatment_pred_cols/n_outcome_true_cols off a compiled model's CausalLoss instead of hardcoding them at each call site -- the root cause of a real shape-mismatch bug found in a demo notebook, and the same pattern that caused the predict_ute_* bug above.
• models.get_propensity_state_conditional_means(model): since the propensity head is no longer a single pypress.keras.layers.PredictiveStateMeans layer, reconstructs the same per-state constants (post-sigmoid) from the model's propensity_logit_state_ layers -- the replacement for the old model.layers[-2].state_conditional_means sanity check.
• metrics.uniform_entropy_gen: mean Shannon entropy of predictive state weights across a batch.

Other fixes

• BiasOnly.get_config() dropped units (silently reverting to 1 on reload) and stored bias_regularizer raw instead of via tf.keras.regularizers.serialize/deserialize.
• The three bootstrap_* functions seeded their resampling RNG inconsistently (RandomState(0) vs RandomState(n_samples)); all three now take an explicit random_state: int = 0.
• README.md's and docs/inference.md's code examples called a nonexistent inference.predict_ate(...) and called utils.split_y_pred(preds) with the wrong arity/unpacking -- anyone copy-pasting the README example hit an immediate error. Fixed and verified end-to-end.
• recommended_callbacks renamed to get_default_callbacks; monitor is now a required argument instead of defaulting to "val_loss".
• Bumped pypress dependency v0.0.6 -> v0.2.3; Uniform regularizer's l1 penalty argument renamed to l2 to track pypress's L1 -> L2 entropy penalty change.
• Added CHANGELOG.md, backfilling v0.0.1-v0.0.13 and documenting v0.1.0.

Version bumped to 0.1.0 in pyproject.toml (breaking propensity-head shape change). Re-ran and fixed the tracked demo/example notebooks against the new API