fix: warn when pyro posterior params fall back to uniform

[nceglia]

Summary

• _ensure_pyro_posterior_params re-initialized q_p_ct_raw to a uniform 1/P simplex whenever the Pyro param store lacked it — silently. This masked failed or missing param-store loads, so downstream posterior metrics ran on an uninformative prior with no signal to the caller.
• Emit a RuntimeWarning before the fallback, naming the affected metrics (joint_distribution_posterior, phenotypic/clonotypic entropy, mutual information) and the likely cause (param store not persisted/restored).
• The early-return path (when q_p_ct_raw is already present) stays silent — no behavior change on the happy path.

Why

The fallback originally existed because PyTorch >= 2.6 weights_only=True made pyro.get_param_store().load() throw silently; the caught exception left metrics running on a uniform prior with no error. With the torch>=2.6 load fixed, this fallback should now only fire on a genuinely empty/missing store — and a re-initialized (uniform) posterior is almost never what the caller wanted, so it should be loud.

Tests

• NUMBA_CACHE_DIR=... MPLCONFIGDIR=... pytest tests/test_pyro_params.py -v -> 2 passed
• Full suite -> 23 passed

New tests/test_pyro_params.py: warns on empty store (and still populates the param so callers can proceed), silent when already present. Both snapshot/restore the global Pyro param store to avoid cross-test leakage.

Closes Notion #24.

🤖 Generated with Claude Code