Translate the base into the destination when using functional record update

[pcwalton]

We have two ways to codegen functional record update:

1. Translate the supplied fields in place, translate the base to scratch, and copy over the delta.
2. Translate the supplied fields to scratch, translate the base in place, and copy over the delta.

Currently we take approach (1). This is not as efficient as approach (2) in most cases, since the whole point of functional record update is usually to update a small number of fields. This would save a few memcpys that show up in I/O.