Experimental FMA support

[dfsek]

This PR adds experimental FMA support as well as several other general optimisations. FMA support can be enabled by setting the JVM option -Dparalithic.optimisation.fma=true.

It is disabled by default because I have not thought of a good way to check if the platform supports intrinsic FMA, and if it does not, FMA is much, much slower than a standard multiply-add.

I also redid the performance tests in JMH to get better metrics.