There is an unidentified condition that prevents udevd calling the
BLKRRPART ioctl after a partition table is written. It looks like
either the kernel or udevd drops device change events if they are
too closely spaced in time. So, in the case where we expect udevd
to call BLKRRPART, check /proc/partitions to make sure it has done
so. Arbitrarily try 5 times, 100ms apart, before giving up and
informing the kernel ourselves.

diskdrake allows the user to clear all partitions even when some of those
partitions are currently mounted. partition_table::dos::need_to_tell_kernel()
must return true in this case, as the automatic reread of the partition table
triggered by udevd will fail.

This avoids unnecessary udev events, which in some cases were causing
udevd to trigger a kernel partition table reread, causing the kernel
to get out of sync with drakx.

In order to do so, the libparted bindings are enhanced so that we
manipulate a 'ped_disk' object:
- disk_open() returns such an object
- set_disk_type() is removed
- disk_commit() writes back the partition table and tell the kernel
  about it
  if we fails to write the partition table, we let
  partition_table::write() fire an exceptionc
  if we only have succeed in writing the partition table, we let
  partition_table::write() try harder (umounting partitions before
  trying againt to tell the kernel again)
- need_to_tell_kernel() is overrided

The last change is needed because we now tell the kernel about the new
partition layout in partition_table::gpt::write() when calling
disk_commit() while previously we were deferring that to
partition_table::write() which after having called the gpt's write()
calls tell_kernel() -> c::tell_kernel_to_reread_partition_table()

So we must tell partition_table::write() that telling the kernel to
reread the partition_table is not needed if we already succeeded in
that.

Now we force the kernel to reread the partition table when we initialise
it, there's no need to also inform it that we've deleted all the existing
partitions.