The automatic rescan of the partition table is triggered by udevd. The
udev rule that causes this is not present on the cut-down system that
runs the classic installer.

To minimise the changes this close to mga6 release (and until we fix
the bug in partition auto-allocation that mistakenly creates BIOS boot
partitions non-GPT disks), reuse the BIOS_GRUB flag for flagging Empty
partitions, as the real partition ID (0x00) is used to flag free space.

When no partitions on a DOS-partitioned disk are mounted, the kernel
automatically rescans the partition table when it is written to disk.
We shouldn't then try to update the kernel's view of the partition
table, as the list of deltas we have recorded is relative to the
previous state of the partition table, not the newly rescanned state.

The behaviour for other partition table types is unchanged.

v2 (tvignaud): just make base class assume the kernel doesn't reread, only mbr
subclass overrides need_to_tell_kernel() in order to be smarter

When no partitions on a DOS-partitioned disk are mounted, the kernel
automatically rescans the partition table when the file handle to the
raw device is released. Currently the code opens and closes the raw
device when writing the primary partition table and when writing each
extended partition table segment. As the extended partition table
segments form a linked list, this allows the kernel to get in and
rescan the table when the list is not in a coherent state. This patch
changes the code to open the raw device before writing the primary
partition table and to close it only after writing the last extended
partition table segment.

The behaviour for other partition table types is unchanged.

v2 (tvignaud): simplify by moving copies of noop funcs into the base class

same rationale as in commit 767048570e8c44061cb0d6faf689698d3313870c for
mga#18666

this wasn't an issue as we later round partition on 1MB boundary but
it's still cleaner/safer...

in order to prevent overlapping errors when adding partitions:
(b/c libparted doesn't know that diskdrake already removed the
partitions in its memory view)
"failed to add partition #1 on /dev/vda"
"Error: Can't have overlapping partitions."
"add_partition failed"

Resolves: mga#18666, mga#17796

basically reverting: commit 36c2fc004bf445ff784bde423ba7aa27885539f8