This was designed in attempt to avoid long buffer scans in bidi_find_paragraph_start.
The same stuff may be used to implement simple per-buffer charpos <-> bytepos cache.
Caching the result of bidi_find_paragraph_start may improve the speed of backward
scrolling/movement (I've seen ~6x speedup for 60M buffer with average string of 5K
characters). Comments are highly appreciated.

Dmitry