This drops the unnecessary bdrv_truncate() from, and also improves,
cluster allocation code path.

Before, when we need a new cluster, get_cluster_offset truncates the
image to bdrv_getlength() + cluster_size, and returns the offset of
added area, i.e. the image length before truncating.

This is not efficient, so it's now rewritten as:

  - Save the extent file length when opening.

  - When allocating cluster, use the saved length as cluster offset.

  - Don't truncate image, because we'll anyway write data there: just
    write any data at the EOF position, in descending priority:

    * New user data (cluster allocation happens in a write request).

    * Filling data in the beginning and/or ending of the new cluster, if
      not covered by user data: either backing file content (COW), or
      zero for standalone images.

One major benifit of this change is, on host mounted NFS images, even
over a fast network, ftruncate is slow (see the example below). This
change significantly speeds up cluster allocation. Comparing by
converting a cirros image (296M) to VMDK on an NFS mount point, over
1Gbe LAN:

    $ time qemu-img convert cirros-0.3.1.img /mnt/a.raw -O vmdk

    Before:
        real    0m21.796s
        user    0m0.130s
        sys     0m0.483s

    After:
        real    0m2.017s
        user    0m0.047s
        sys     0m0.190s

We also get rid of unchecked bdrv_getlength() and bdrv_truncate(), and
get a little more documentation in function comments.

Tested that this passes qemu-iotests for all VMDK subformats.

Signed-off-by: Fam Zheng <famz@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>