There are three major components left to implement in `sais` before it can be
realistically called "done."
These are emphatically in no particular order. In fact, I do not know in which
order I should tackle them.
1. Figue out the story with sentinels. Do we really need them in the string?
It seems trivial enough to remove them and just special case an the index
equal to the length of the string. But...
2. Implement generalized suffix arrays. The trick with these is that you can't
incorporate suffixes that cross word boundaries. It seems like it should
be possible to do this without sentinels, but sentinels may make the code
much simpler.
Generalized suffix arrays also require a lot more memory.
But generalized suffix arrays are *really* useful.
3. Reduce allocation. We really need to focus on not creating `Vec`s of size
equal to the number of input characters. We're doing a lot of that now.
Similarly, we should try to figure out how to *not* have `wstrs` and
`wstrs_sorted` in memory at the same time. Maybe induced sorting will help
with this?
A paper describes a trick to keep memory requirements very low by reusing
the suffix array. I think we can probably ignore this optimization for now,
but we still need to reduce allocation.
4. Profile. There are undoubtedly places where I've been careless and things
could be done faster. e.g., Every character iterator is running a
method like `v.char()` or `v.idx_char()`. Are these getting inlined?
I hope so.
5. Do we pay a cost for the generic `Text` trait? I thought a lot about how
to do this without a `Text` trait, and I couldn't come up with anything
workable. We really need to write code that is generic over iterators of
characters...