iomap write fixes

[Christoph Hellwig]

These have already been discussed in a lkml thread for a while, so
I wanted to formally send them out.  The problem was that as part of
a AIM7 regression we saw that we spend a lot of time looking up
extent records.  I turns out that the current delayed write path
is horribly inefficient, and we can generall do with a single btree
lookup for the write path.  Also xfs_inode_set_eofblocks_tag turns
out to trash the cache quite bad, so add a trivial flag to avoid
it if we can.

Compared to the git tree under discussion only one changelog has
been updated, and I've added a copyright statement to xfs_iomap.c
now that it has a lot of my code.