Use gcc section attribute to keep less frequently used functions
away from main code.

Make sure the writer thread starts and stops when we expect it to.

And mdb_env_copyfd2(). Perform compaction on the copy. Trims out
freed pages and renumbers data pages in sequential order. This is
more CPU-intensive since it copies and modifies data pages.

Don't try to write past end of file

Do not overwrite loose pages' mp_pgno with mp_next.
mdb_page_flush() will use mp_pgno to rebuild dirty_list.

Try to exit cleanly to avoid leaving stale readers. Not
a critical issue since mdb_reader_check can take care of
them, but still cleaner.

Require that the page is dirty in the current txn,
not a parent txn.

Rename MDB -> LMDB
Integrate tool manpages

Just use srand/rand, we don't care about the quality
of the random numbers, we just want some data.

Don't try this when working on the FreeDB. Again,
too much hassle to unkink the recursions...

Only loosen src page if it's actually dirty.
Refactor page refs in page_merge.

Pages that were dirtied and deleted in the same txn should be
reused, instead of consuming freeDB pages.

Similar to prev commit for mdb_node_move, the call of
mdb_page_serch_lowest() was leaving csrc unusable.

Was leaving stale info in dst cursor

Use num*20, chosen from empirical testing

Was skipping the overwrite in some cases, if the cmp function
said the new and old already matched. We should always perform
the overwrite anyway, since the cmp function may only be doing
a prefix compare and the data may actually be different.