Break down u256 into u128 limbs

[DavePearce]

The goal here is to break down columns with type u256 into arrays of two u128 limbs. A key question is where in the pipeline should this happen? Should it be when going from HIR to MIR, or from MIR to AIR, or something more unusual like from MIR to MIR?

Logically, it feels like MIR -> AIR is the right place. The downside of this, however, is that AIR would then be the only place where field arithmetic was performed. It could be that we could low from big.Int to fr.Element at the MIR level perhaps? That would certainly be interesting.