Previously a read transaction could be held open by each fork
indefinitely.  That was done for better speed, but it had a downside
of keeping old pages alive and potentially reading only old data,
until some writes were attempted by that fork.

Now kr_cache_ provides explicit API for suitable points where to break
transactions, reusing the _sync command.  On LMDB side the read-only
transaction is only reset and later renewed, supposedly giving better
performance than aborting (see LMDB docs on reset+renew).

Performance: preliminary testing with two forks, resperf on comcast
query-set shows no noticeable difference in peak QPS.