JIT witgen code generation performance through occurrence maps

[chriseth]

For large number of identities (arith256_memory), it takes a long time to go through the full list of identities.

Instead, we should create a map from all variables to the identities they occur in. Once we get an update about the information of a variable, we only take a look at the identities the variable occurs in.

We could still try to always try to process the identities in source order, but that should not be a problem:

We can maintain a sorted queue or something:

// TODO maybe better to use a BinaryHeap
let mut identities_to_process: BTreeMap<usize, &Identity<T>> = all_identities(); // the usize is the index
while let Some((index, id)) = identities_to_process.pop_first() {
  // We don't need `progress` any more, since "progress" just means "at least one var was updated"
  let updated_vars = process_identity(id);
  identities_to_process.extend(updated_vars.iter().flat_map(|v| &occurrence[v]));
}