One population-path interpreter over representation-specific batch phases.
Value phases may complete synchronously and contain fused numeric loops;
trace phases may suspend. Both use the same weighting/selection/retention.

Particle populations on θ vectors: the shared engine of the samplers that
move a model that is one block (`foerster.block`) without worlds.

A model that is one latent block site and nothing else is, for inference,
its θ: its density is the block's, and the program around it only computes
the result from θ. Such a model runs here on plain vectors, and a world is
made only at the end, by one replay per distinct final θ (`write-back-all`),
so every result is still computed by the program itself.

`theta-smc` is an SMC sampler over a sequence of targets on θ, given by a
path: tempering from the block's draw density to its target is one path
(`foerster.inference.tempering`); data arriving in batches or a hierarchy of model
levels are others. A path decides the next position from the population,
gives each particle's incremental log weight to it, and the log density the
moves at that position leave invariant. Resampling, the moves and the
evidence are the engine's.

Randomness is keyed by what is drawn (`foerster.random`): the resampling
at step k and the moves of chain i at step k read their own streams, under
keys the caller names, so a sampler moved onto this engine keeps its
draws.