lsqisotonic(x,y,w)

This is a more efficient implementation of the private "lsqisotonic" function for the "mdscale" function in the statistics toolbox, which performs non-metric multi-dimensional scaling.
In non-metric multi-dimensional scaling, one of the steps is to find a monotone regression of the dissimilarities that has the least squared error, which is the what the function "lsqisotonic" does. During this procedure, it is necessary to repeatedly merge adjacent monotone blocks.

The original implementation uses an iterative procedure, which has a complexity of O(n) per iteration, and O(n^2) overall. When n, the number of data points, is large (say, > 1000), the "lsqisotonic" function may be very slow. I replaced it with a non-iterative procedure that has an overall complexity of O(n).