`--merge` should memoize the results for performance

--merge currently takes a long time, needlessly.

Instead of iterating over each locale in turn and merging in the ancestor locales each time:

en-CA -> en -> root
en-GB -> en -> root
en-US -> en -> root

All of these use the en -> root, which is the same data, so there is no need to recompute those for each child locale.

Instead, you could iterate over the graph of locales breadth-first starting at the root locale, then cache the results for use in the other locales.

(Of course, this might not be worth doing as the whole concept of --merge is likely to change. I just wanted to capture this potential optimization here)

Start by locating the --merge implementation and tracing how ancestor locales are merged for each locale. Compare the current traversal with the proposed breadth-first approach from root, then verify that shared ancestor data is computed once and the merged output remains unchanged. The issue names no specific files or tests.