Instead of modifying the accumulated expressions in-place, we now build a set
of expressions that are known to be zero, and then consult that set on the fly
when converting the expression data for FFI.
This will be necessary when moving mappings and expression data into function
coverage info, which can't be mutated during codegen.