improvements to neighbor joining

[gregcaporaso]

@ElBrogrammer and I discussed this while working on #420.

• [ ] should investigate building TreeNodes as the algorithm proceeds, rather than newick string, this will likely be more efficient (but maybe not if the user just wants a newick string back, since it'd be a lot of TreeNode instantiation)
• [ ] some functionality should be ported to methods of the DistanceMatrix, including:
  • [ ] min, max (see #228): the current _lowest_index method
  • [ ] collapsing a pair of values: the current _compute_collapsed_dm and _otu_to_new_node methods
  • [ ] probably should rename _otu_to_new_node since technically it's computing the distance from a node (either internal or a tip) to the new node

Anything else @ElBrogrammer?

Depends on:

• [ ] TreenNode Refactor (#809)
• [ ] DistanceMatrix Improvement (#824)

[jairideout]

Maybe _compute_q? Though I'm not sure how generalizable/useful that'd be as a DistanceMatrix method.

[jairideout]

Would be great to add a DistanceMatrix.nj method (the actual implementation could stay where it's at).

[wasade]

TreeNode instantiation isn't that bad for large trees, and if the primary use case is to get a tree back, then I think it would be good to construct the tree on the fly and let the developer/user use TreeNode.to_newick if they need a str representation.

Agree re: changing OTU to node also because this isn't specific to OTUs