qEI mvndst accuracy docs

[denizokt]

quoting eliu:

it'd be nice to have some docs (could even be in a separate document somewhere) explaining where 1e-5 and 2000*u.size came from.

Like mvndst docs recommends 1000 * u.size, that python wrapper we looked at uses 10000 * u.size, is 20k just to be safe? Is this ever enough?

And with 1e-5, how does that translate into the accuracy of the overall qEI computation? Is it a strong impact or are we already seeing variations only in like 10th decimal place?

[suntzu86]

also from the original qEI review:

@sc932 should we be logging error or flag when accuracy was insufficient? Should we attempt to call the integrator again with more iterations?

in particular, we always pass back the result of mvndst (return out[1] line 276 of python_version/expected_improvement.py). that method provide some self-diagnostics that we should at least understand (if not log or react to).