Starts near the end of the file and reads forward, looking for a
complete snapshot.  If it doesn't find one, move farther from the end
and repeat.

XXX This approach does N**2 reads, where N is the number of times we
need to move further back and retry.  When we move backwards, we
double the distance from the end of the file.

This processes the whole file from the beginning collecting info about
active / blocked transactions.  This version is preliminary because it
needs to start at the end and backup until it has enough information.