xfs_copy: actually do directio writes to block devices

xfs_copy: actually do directio writes to block devices

Not sure why block device targets don't get O_DIRECT in !buffered mode,
but it's misleading when the copy completes instantly only to stall
forever due to fsync-on-close.  Adjust the "write last sector" code to
allocate a properly aligned buffer.

In removing the onstack buffer for EOD writes, this also corrects the
buffer being larger than necessary -- the old code declared an array of
32768 pointers, whereas all we really need is an aligned 32768-byte
buffer.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>