Implement EnsembleServer.marginals()

[fritzo]

There are two types of servers in serving.py: the TreeCatServer for single latent hypotheses, and EnsembleServer for ensembles of multiple latent hypotheses (typically model_ensemble_size = 8).

The .marginals() method of serves should compute marginal probabilities P(Z_v=z|data) for each vertex v, where Z_v is the observed multinomial variable at vertex v, and data is a partially-observed row of conditioning data. This method is currently implemented for individual TreeCatServer objects, but not for aggregate EnsembleServer objects.

To implement .marginals() for EnsembleServers, we need to compute the individual self._ensemble[i].mean(data), and then return a weighted average of these w.r.t. self._ensemble[i].logprob(data).

Note that data here is specified in the internal format of an [N, R]-shaped ragged array, with blocks defined by ragged_index.