This will allow us to store the id
in the struct itself.

There is no need for IndexMap here,
we don't iterate or care about indices.

We used to have the concept of "resetting"
a tracked struct, which removed all instances.
We don't have that concept anymore.
Therefore we don't need to record a read on the
tracked struct table itself.

This is a step towards refactoring `fetch`
to return the memo itself and not the value.

Change the core operation to be `refresh_memo`.
This is a nice refactoring on its own but it
will also support the refactoring of how
we manage accumulator values.

We used to store values in a central map,
but now each memo has an `AccumulatorMap`
that maps accumulated values (if any).

The primary goals of this change are

* forward compatible with speculative execution
  because it puts more data into tables;
* step towards a refactoring to stop tracking
  outputs in the same array as inputs and thus
  to simplify how we do versioning. We will no
  longer need to walk the function's outputs
  and refresh their versions and so forth because
  they are stored in the function memo and so
  they get refreshed automatically when the memo
  is refreshed.