Add synthetic-EHR generative evaluation metrics

[chufangao]

Summary

Adds pyhealth/metrics/generative/, a subpackage for evaluating synthetic
EHR data along three axes — privacy, utility, and statistical
fidelity.

• privacy.py — calc_nnaar (Nearest Neighbor Adversarial Accuracy
  Risk), calc_membership_inference (membership inference attack), and
  compute_discriminator_privacy (real-vs-synthetic discriminator score).
• utility.py — compute_mle (machine learning efficacy, TRTR vs TSTR)
  and compute_prevalence_metrics (code-prevalence similarity: R², Pearson,
  RMSE).
• utils.py — shared data prep, a self-contained LSTM classifier, and a
  random-forest baseline.
• evaluate_synthetic_ehr() — convenience orchestrator that runs the
  full suite and returns one merged {metric: (mean, std)} dict.

The metrics are ported from a standalone evaluation script. The
MIMIC-specific data-loading/CLI glue is dropped so the functions work on any
flat EHR dataframe (one row per patient/visit/code event). Public functions
are re-exported from pyhealth.metrics.

Cleanups applied during the port

• logging instead of bare print calls.
• Fixed a latent CUDA crash in the LSTM eval loop (.cpu().numpy()).
• Replaced scipy.stats.pearsonr with numpy.corrcoef to avoid an
  undeclared scipy dependency.
• Input dataframes are copied instead of mutated in place.
• Google-style docstrings, type hints, PEP8 (≤88 chars).

Tests

tests/core/test_generative_metrics.py — 18 unittest cases, all passing:

• 13 functional tests covering each metric and the orchestrator
  (lstm + rf modes, argument validation).

• 5 behavioral tests (TestMetricsBehavior) that verify each metric
  responds sensibly across three synthetic datasets — an exact copy of the
  training data, a similar set (~15% of codes perturbed), and a different
  set (disjoint code vocabulary):

  Metric	Verified behavior
  Prevalence	RMSE 0 → 0.03 → 0.26; exact copy → RMSE 0, R²/Pearson = 1
  NNAAR	Flags memorization: 1.0 → 0.1 → 0.0
  Membership inference	Attack accuracy 1.0 → 0.94 → 0.46 (chance for unrelated data)
  Discriminator privacy	Disjoint-vocabulary data trivially flagged; real-derived data is not
  MLE (utility)	Exact copy reproduces real utility exactly; ratio degrades 1.0 → 0.98 → 0.81

Docs

Added docs/api/metrics/pyhealth.metrics.generative.rst and a toctree entry
in docs/api/metrics.rst.

Notes

The discriminator-privacy score is degenerate for exact copies (the model
predicts a constant on identical features, so the score reflects test-split
balance rather than 0.5). The behavioral test asserts the robust direction —
disjoint synthetic data is cleanly flagged while real-derived data is not.