Add sample size-based ML/REML estimators

[tyarkoni]

Adds a new SampleSizeBasedLikelihoodEstimator class that produces maximum-likelihood random-effects estimates for datasets that have estimates and sample sizes but no variances. This is maybe the modal situation with fMRI data, so it's a nice thing to have.

The approach is based on a simple idea in this paper (though I'm sure it's been suggested before in other places), which is to replace the standard log-likelihood function for the random-effects model with an analogous LL based not only beta and tau^2, but also a sigma^2 parameter that's scaled by the (known) inverse sample size. I.e., instead of study estimates being distributed with a variance of tau^2 + v (where v is the vector of observed sampling variances for the available studies), we assume they're distributed with a variance of tau^2 + sigma^2/n (where n is the vector of known sample sizes, and sigma^2 is an additional parameter estimated from the data).

I haven't done any systematic investigation of the performance of this estimator, but the above paper suggests it works reasonably well in many cases, and if nothing else this seems like an improvement over just fitting a GLM to the estimates directly (i.e., ignoring sampling variance/sample size entirely).