Review the R parsimony benchmarking and less horriblyify the code

[jeromekelleher]

The R_implementations.R is my hacky attempt to get something to work in R. However, I'm substantially less confident that we're correctly measuring things there, since I'm a crap R programmer.

Some points to consider:

• Is there any way to stop phangorm from finding the unique site patterns? Is what I'm doing correct? We're reporting the average per-site time, so we need to know how many sites are actually being processed.
• Is the implementation really vectorised, written in C etc? It's a really strong point if we can say that (fairly obscure) C code embedded in the R interpreter is slower than JIT compiled Python, so it would be good if someone with stronger R skills could take a look.

@petrelharp, do I hear you volunteering for this 😉

[jeromekelleher]

Might be worth looking at Castor (#14) while in there to see if it supports Sankoff/Hartigan parsimony?