We're spending huge amounts of time on lock contention during
end_io processing because we unconditionally assume we are overwriting
an existing extent in the file for each IO.

This checks to see if we are outside i_size, and if so, it uses a
less expensive readonly search of the btree to look for existing
extents.
Signed-off-by: NChris Mason <chris.mason@oracle.com>