First bug is, we were passing as count to th_brk, the UTF-8
length instead of TIS length.  Ouch!  I'm not sure how this
was never caught...

The other one was, break-thai was possibly marking a position
is_line_break when break.c has already set it to is_char_break=FALSE.
This broke the invariant that if a position is line-break, then it
must be char-break as well.  This, in turn was hitting assertions
in certain conditions.  Hit it with this for example:

$ ./pango-view --text 'ส์" (S' --width 43 --font 156px

Note that in a correct world the Latin part of that string should
not reach break-thai.c at all, but that's not how pango-layout.c
breaks right now.  See comment before pango_break() call in
pango-layout.c.