Until now, searching used to be done using a single JS object indexed by
word:type strings. That means, to search for a noun we'd do two searches
into this object: query:n and query:n:pr. And to search for a verb, we'd
do many searches (query:v:in, query:v:tr, ...). And worse, to search for
anything else, we'd just linear search through the list. This clearly is
suboptimal, but was grown this way for historical reasons.

This commit is a big refactor to improve this. Now, the word database is
simply an array. A word can be indexed in two ways: (1) with their index
in the array, and (2) with a word/type pair like before, but now we do a
preprocessing step to create an object that maps words (independently of
their type) to a list of array indexes. This way we can answer questions
like finding a word in essentially O(1) time instead of O(n).

A secondary benefit (actually the one because of which I finally decided
to get started with this big refactor) is that now that we are not using
the "na'vi" field for searching anymore, it becomes possible to put some
extra data in the "na'vi" field. Specifically Reykunyu has pronunciation
and infix data, but these are separate fields from the "na'vi" field, so
it is not (easily) possible to show infix dots or underline the stressed
syllable in the lemma itself. This commit allows indicating the syllable
boundaries with "/" and the stressed syllable with "[...]"; this is then
shown in the lemma display, without affecting searches. Also, we can now
put ù in the field so that we can automatically create Reef Na'vi forms,
if the user wants to see RN. As such, this commit also adds a (as of yet
still non-functional) setting to the UI to switch between dialect modes:
FN, RN, or a combined mode.

Right now all of this is very experimental. Many features (e.g., the all
words page and the editor) don't work. I still have to reconsider how to
exactly deal with the FN/RN distinction. For example, in word links what
dialect should be used? I don't want to have to write [[to]/rùk:n] for a
word link to [toruk:n]. So, should word links just use FN spelling?

When this all works, this should address #70, #77, and #105. Eventually,
it could also enable addressing #36, #42, and #57.

Fixes #135.

Word data that is independent of the user query is now processed when
the dictionary file is loaded, on Reykunyu startup, instead of every
time the user does a query.

Fixes #138.

Move word data processing out of the postprocessing step

And show only the pronunciation IPA of the selected dialect.

Fixes #137.

This (ridiculously complicated) code implements the 'awkx → 'awgìl
pattern in RN.

For example, tì'i'a; originally only one of the two 's would be removed.

Alternative title: Wllìm figures out how to use lookahead.

Because the dialect was only set after the on-page-load search was
triggered, the search would always use the combined dialect.

This way, clicking them doesn't result in a page reload.

Fixes #148. The proper noun doesn't properly get capitalized yet, but
that's another issue.

The problem was an overzealous parsing rule. If a noun ends on -g (after
other affixes, such as -ìl, already were removed) then the rule replaced
the -g by -kx, and because that's the only possible option, I didn't
even make it bother to check the variant with -g. That doesn't work
however for nouns ending in -ng, because it would become -nkx, which
obviously doesn't make sense.

Fixed this by just also checking the unchanged variant (exactly how all
the other rules work).

Tatlam ke lengu oe tsùlfätu lì'fyaye wione.