close[t:5128] updating a descriptor now has two parts: user API that passes a FT and descriptor, and a low layer API that takes those two plus an explicit descriptor. we use the low layer one for deserialization code that wants to upgrade a descriptor for an FT that is not fully opened yet (and thus whose FT does not have a valid cachefile yet). this seems to be a clean enough solution, since it helps some upgrade code do non-standard stuff.

most importantly, loader-stress-test3 no longer fails by segfaulting on ft->cf->fd


git-svn-id: file:///svn/toku/tokudb@44839 c7de825b-a66e-492c-adef-691d508d4ae1