of the benchmark (parsing a repeated message, where each message had 5 "small"
enum values and 5 "large" enum values) from ~39s to ~11s.

There is a small memory cost per enum used, but I expect this to be trivial compared
with other per-type costs.

Fixes issue 97.