remove write lock from _migrateStartCloneTransaction

[leifwalsh]

Currently, the beginning of a migration takes a write lock to start the migration. If there is a long running read operation, that can introduce a very bad stall. Instead of this, we'll use a range lock in the lock tree over the chunk. This introduces the invariant that clients must take row locks before running the oplogging code (which is currently where we log into the migratelog too), so that this lock will be a valid way to make sure nobody's writing to the chunk while we're initializing the migratelog.

[leifwalsh]

Consider using locktree locks instead, need to have a pow-wow with @zkasheff