Low latency write-only transactions

[craigfe]

As discussed in the meeting, there is a possible strategy for having a period of many successive writes to an index that are not readable until after all writes are completed:

offload writes to a set of rolling log files, then at the end of the transaction we merge-sort each log file into the data file and resume the standard operating mode with a single log file.

CC @pascutto, @ioana, @samoht.