Gi rewrite

[ljwolf]

The G_Local statistic currently assumes that all weights are equal in its __crand() method. This could be in the case where input weights are binary, or in the case where input weights are row standardized, but that row standardization has to be of equally-weighted input. This isn't valid when input weights are row-standardized, but not constant (like, in perimeter-weighted contiguity or any kernel weighting). While the vast majority of users use binary G_Local, this would be good to fix.

So, to simplify the calc method and correct this in the __crand implementation, @sjsrey and I have drawn up a roadmap of the desired new behavior:

• change the star=False arg to star=None
• use the w however it is provided if star=None. This means that, by default, the statistic will:
  • detect if there is a diagonal value in w.sparse
  • if so, treat it as Gi* and otherwise compute Gi.
• if star=False, the input w will have its diagonal zeroed. If transform='R', then the weights will be re-transformed after over-writing the diagonal.
• if star=True, then we need some more information:
  • ifw has a diagonal, use it.
  • if w has no diagonal, then examine the inputted weight's transform:
    • if the weights are binary, add one to diagonal
    • otherwise, warn that the statistic is going to be the gi, not gi-star and give directions on how to update the w with a self-weight for gi-star.
• if star is a float, use that as self weight if the w doesn't have it