runtime: Improve the map implementation.

runtime: Improve the map implementation.

New version is somewhat faster and use somewhat less memory.
Mapping is done in two stages:
1) mapping from the hash value to a dense
   range of cell IDs.
2) mapping from dense range of cell IDs to
   key/value pairs.
Keeping the key/value pairs in a separate,
growable-array structure makes the iteration
semantics easy.  The implementation of #1 can
then ignore iteration and just go for fast/dense
encoding.

Experiments put lookup at ~20% faster and insert
at about 50% faster.  Also uses slightly less
memory.

First bug-free version.