Add sample.posterior.resimulate flag for posterior predictive checks

[oleg-savchenko]

After sample_posterior(), only latent (estimator) node values are saved by default. When resimulate: true, this commit runs each posterior sample through the forward model and saves all node values together, enabling posterior predictive checks without a separate script.

New DeployedGraph.resimulate_posterior() method encapsulates the pattern already used inline in the adaptive training loop. Called from both sample_mode() (falcon sample posterior) and launch_mode() (end-of-training posterior sampling). _save_samples() default key selection extended to cover all nodes when resimulate=True.

Default behaviour (resimulate: false) is unchanged.

Motivation

falcon sample posterior currently saves only the latent inference nodes
(e.g. delta_ic). To do a posterior predictive check — draw
theta ~ posterior, then simulate x_pred ~ p(x | theta) and compare
to the observation — users currently have to write their own forward-pass
script outside of falcon.

This PR adds sample.posterior.resimulate: true (default false, no
behaviour change), which after drawing posterior samples automatically
runs them through the forward model and saves all node values together.

How it works

The pattern mirrors what the adaptive training loop already does internally
(draw proposal → forward-simulate for buffer). The logic is extracted into
a new DeployedGraph.resimulate_posterior() method so it can be reused from
both sample_mode and launch_mode.

Changes

• falcon/core/deployed_graph.py: new resimulate_posterior() method
• falcon/cli.py: call it in sample_mode() and launch_mode() when
  sample.posterior.resimulate=true; update _save_samples() to include
  all node keys when resimulating
• docs/configuration.md: document resimulate, exclude_keys, add_keys

Usage

sample:
  posterior:
    n: 100
    resimulate: true

[cweniger]

I suspect this is now obsolete give that you can use 'sample ppd'. The advantage of not mixing posterior and ppd sample paths is that often one might want to have many more (cheap) posterior samples than (expensive) ppd samples. Thoughts?