Replace BTreeSets used in MergeExpressionsOptimizer with an enum

[TomAFrench]

In merge_expression.rs, we currently track all of the opcodes which reference a particular witness index in a

BTreeMap<Witness, BTreeSet<usize>>

As we don't need to track these sets past 2 opcodes referencing a witness, this would be more efficient if we were to instead use

enum ReferencingOpcodes {
    One(usize),
    Two(usize, usize),
    Many,
}

BTreeMap<Witness, ReferencingOpcodes>

This avoids needing to track witnesses other than the ones which can potentially be optimized out in future. This means that even in the case where a million opcodes reference a single witness then we can represent this with the memory equivalent of 2 usizes.