More optimisation steps for topology

[johnlees]

Don't think we ever tried adding NNI and SPR, but would probably be sensible. The radius approach from MAPLE for SPR seems sensible too

[jhellewell14]

Added NNI here: fbbe614

At the moment it just performs a NNI move. I need to put in some kind of structure where there is a "stable" and a "working" tree. The working tree is a copy of the stable tree, I then do moves on it, calculate the new likelihood of the working tree and "accept" them by writing this back on to the stable tree.

I do something similar for hillclimbing:

trees_rs/src/hillclimb.rs

Line 38 in fbbe614

impl<T: RateMatrix> Tree<T> {

But I only save the stable and working tree vectors. For the hillclimbing this is fine because each random move starts from the last random move. For NNI each move will need to start from the stable tree rather than the previous tree.

[johnlees]

Also if you wanted to parallelise this then I think you could keep the stable tree and pass it by ref, then in each new thread start by copying it, do the NNI move, then in the main thread just keep the one that has the highest likelihood

[johnlees]

It would be quite neat in the rayon iterator format because you could just run max on the function return