fix: strip embedded penalty terms from causal_loss_metric_gen (v0.1.1)

[gmgeorg]

Summary

• causal_loss_metric_gen reused the exact outcome_loss/treatment_loss instances it was given, so any penalty folded into either loss's own call() (e.g. a within-state balance penalty) would leak into the metric used for EarlyStopping/ReduceLROnPlateau/checkpoint selection and any hyperparameter-tuning objective built on it.
• Adds an explicit penalty_free() contract to OutcomeLoss/TreatmentLoss: each reconstructs itself from a hand-written list of its own known-safe constructor args (never by scanning attribute names). A future penalty either defaults to "off" automatically or raises loudly until the subclass explicitly overrides penalty_free() — it can never leak through unnoticed. causal_loss_metric_gen now calls it on both losses before wrapping them.
• Bumps version to 0.1.1, adds a CHANGELOG.md entry, and a dev-docs/ writeup with the full rationale (local/gitignored, per repo convention).

Test plan

• pytest pypsps/tests/test_losses.py pypsps/tests/test_models.py pypsps/tests/test_metrics.py — 26 passed
• pytest pypsps/tests/ (full suite) — 67 passed
• New regression tests: penalty-free identity when no penalty declared, penalty zeroed without mutating the live loss, fail-loud for an undeclared required penalty arg, and an end-to-end check that causal_loss_metric_gen's output matches the unpenalized likelihood, not the penalized one.