Improve performance for large sample sizes

[longouyang]

No description provided.

[hawkrobe]

Any ideas on how to start on this? Calling viz.marginals on a probmods example with >100,000 samples (and four variables in the joint distribution) takes an order of magnitude more time than inference

[longouyang]

There are a couple of bottlenecks:

1. The projected-out distributions are computed in a concise but inefficient way
2. Density estimation for continuous data is implemented naively. There are various tricks (e.g., FFT and tree-based computation, other stuff) to make this faster but it might not be worth the effort, given that we'll probably want kernel-based aggregators in core webppl anyway (Thoughts on design of aggregators / marginal ERPs webppl#369). Also, if I had to guess, the main bottleneck is probably the projection, not kde.

[hawkrobe]

Thanks for the tips -- I might take a look.