ICU non-utf8 code path leaks memory like there's no tomorrow

In a database with utf8 encoding, this behaves reasonably:

select count(*) from
(select * from generate_series(1,10000000) x
order by x::text collate "en-x-icu") ss;

It eats circa 25MB, not a lot worse than the libc-collation equivalent.
But try it in say LATIN1, and it eats multiple gigabytes.

I believe the reason is that the code paths in varstr_cmp that make
use of icu_to_uchar() have forgotten to free the palloc'd output
of the latter. I have not looked to see where else the users of
that and the reverse function made this mistake.

regards, tom lane