fix tests that use it to not need it

 - break up cachetable_flush_cachefile into more digestable functions,
 - decouple hash_id from filenum
 - break up close_userdata into close_userdata and free_userdata

cachefiles_list class and move some functionality in there.

Isolate mempool and OMT into a new class, bndata.
Remove key from the leafentry.

txn object after it commits or aborts

BUILD_TESTING=Off in the cmake config.

properly after a small append into a large append

fixes #70

timing-dependent (though it still is)

and a new operation in test_stress0 for stress testing coverage

fixes #65

 - gets indexer to run in reverse, that is, start at the end and run to beginning
 - refines locking a bit. An estimate of the position of the hot indexer is stored,
   that is cheap to look at. Threads that use this estimate with a mutex either do
   only a quick comparison or set it to a new value. Threads doing writes (with XXX_multiple calls)
   will check their position with respect to the estimate, and if they see the hot indexer
   is already past where they will modify, they don't grab the more expensive indexer
   lock. For insertion workloads that go to the end of the main dictionary of a table/collection,
   this check should practically always pass.

BDB versions of tests fail to compile because DBT_ARRAY is undeclared.

BDB versions of tests fail to compile because DBT_ARRAY is undeclared.

refs #48, have ft_flush_some_child still do a flush if the child is a leaf node, so garbage collection happens