Implement total_branch_length using numpy ops?

[jeromekelleher]

After #1704 lands we can implement total_branch_length like this:

@property
 def total_branch_length(self):
        nodes = self.preorder()
        parent = self.parent_array[nodes]
        time = self.tree_sequence.tables.nodes.time
        branch_length = time[parent] - time[nodes]
        return float(np.sum(branch_length[parent != NULL]))

I'm sure this'll be faster most of the time, but I guess there could be some cases where overhead of creating the time array leads to a regression. There'll be a bit of memory overhead too, but I doubt this is significant.

Any thoughts on whether we should/shouldn't do things like this?

(In retrospect this should have been a function rather than a property, as it would have been more useful (you could specify the root node you wanted to sum from), and it also violates the "properties are inexpensive" rule. I think my reasoning at the time was that this was something we should be able to keep track of efficiently incrementally, but this isn't actually true under the current definition, which only sums branch length reachable from roots. That ship has sailed anyway, but I guess we should make a note of this in the comments if/when we're updating the method)

[jeromekelleher]

Although, on second thoughts, we can trivially implement this now in C using the preorder function and there's no downside there, so it would be quicker to just do this than to try to benchmark the pros and cons of the numpy version.

This could be a good example to use for illustrating tree operations using numpy, though (#1788)