This should speed up global code completion by avoiding deserializing
preamble declarations to look up names. The tradeoff is memory usage.
Currently the index is fairly naive and may not be much faster, but there's lots
of performance headroom.
These two changes go together because results from the index get copied a couple
of times, so we should avoid it for huge sets.
Also the flag should be -completion-limit, rather than -limit-completion.