Re-implement NCNUM type where possible

[willend]

To facilitate use of (e.g.) 32bit floats in GPU settings, it would be nice to re-implement the original abstraction from KN.

I.e. to switch by simply compiling using
-DMCNUM=float
(An yes, naturally, the default should be double.)

Some (share) chunks of code are likely "bound" to stay mostly double based, but having the option to switch would be good preparation for e.g. a GPU hackathon.