[WIP] EM SDE Distribution

[maedoc]

This was a comment #211 which snowballed into a PR, which ports PyMC3 implementation of Euler-Maruyama scheme for stochastic differential equations. I'll follow up on

• batch vs event shaping (I don't really understand this yet)
• sampling
• tests/examples

but feedback appreciated

[maedoc]

I'm also wondering if I should implement the "non-centered" version of the this, where instead of N(mu, sd), we'd have an auxiliary distribution N(0,1) which is then used in the SDE step. I understand this is better for an HMC sampler, but not sure if the other inference algorithms in Edward would also benefit from that.

[dustinvtran]

Apologies for the delay. Was busy at NIPS.

This seems very useful. If I understand correctly, it's a vectorized approach to handle the time series. The log prob implementation makes sense.