Buffer updates per-transaction rather than per-update. If lots of

Buffer updates per-transaction rather than per-update.  If lots of
updates happen within a single transaction, there was a lot of wasted
encode/decode overhead due to segment merges.  This code buffers
updates in memory and writes out larger level-0 segments.  It only
works when documents are presented in ascending order by docid.
Comparing a test set running 100 documents per transaction, the total
runtime is cut almost in half. (CVS 3751)

FossilOrigin-Name: 0229cba69698ab4b44f8583ef50a87c49422f8ec