Dedupe replacement lists

[Chris Warburton]

Our replacements are lists of lists. Each inner list represents a set of
equivalent names; the first element is the "canonical" one to use, all
the others will be replaced to use this one. We ensure that the first
name is the smallest lexicographically.

These inner lists can end up with quite a few duplicate elements. We
handle this by providing 'new-of' and 'old-of' instead of using bare
'car' and 'cdr', so that 'old-of' can remove any occurrences of the
canonical name. It's still pretty naff though.

One thing we need to avoid is doing too much processing while these
replacement sets are being constructed; for example, sorting the
elements after an insertion. This is because our deduplicating algorithm
is already O(n^3) so we want to avoid any further work in each
iteration.

One possibility is to dedupe as part of the 'wrapper' functions like
'replacements-closure', or even 'normed-and-replacements' (since it's
'normed-and-replacements-inner' which does the O(n^3) work). In the
latter case, this would ensure that the deduping gets cached.