running and t-running covariance

[shabbychef]

This would subsume #36. My idea is for the function to return a matrix where each row is the n(n+1)/2 elements of the lower triangle of the covariance matrix, then n elements of the means, then 1 element of the (weighted) count.
The only acceptable NA mode would be "complete cases".

[shabbychef]

Running correlation would then just use that? Users could get just a single rho out or a running matrix of the n(n-1)/2 strictly lower triangular elements of the correlation matrix.