[lucianogdittgen]

Problem

Since d6c8282 ("feat: refuse to install over a partition which is in use"), a
raw object targeting a whole disk is refused whenever any partition of that
disk is mounted. Boards keeping their bootloader in the gap ahead of the first
partition can no longer install a package updating it. On an am335x with MLO
at 128 KiB and U-Boot at 384 KiB of /dev/mmcblk1, and the rootfs from 1 MiB:

ERRO device failed vaidation: target device "/dev/mmcblk1" is in use, it holds the filesystem mounted at "/" (DeviceInUse { device: "/dev/mmcblk1", mount_point: "/" })

Neither write reaches a byte of the mounted filesystem. Agents up to 2.1.x
install these packages, so a device updated to 2.2.x loses the ability to
update its bootloader.

Changes

1. fix(raw): only refuse a write overlapping a mounted partition. The raw
   handler checks the region it writes (Access::Region). On a whole disk,
   partitions lying entirely outside of it (bounds from sysfs start/size)
   are ignored, along with whatever is stacked on them. These are still
   checked: the disk itself, its holders, overlapping partitions and
   partitions whose bounds cannot be read. Any other target uses the
   existing check. raw-delta keeps the exclusive check, since its written
   length is not known up front.
2. fix(raw): never write past the region checked for the object. The
   region comes from the object metadata, so the write is now bounded to it
   and the install fails if the data goes further. A compressed object
   without required-uncompressed-size has no known region, so it keeps the
   exclusive check and is not bounded.
3. test(fs): cover the region check on a partitioned loop device. Mounts
   a partition placed after a 1 MiB gap and checks that a write to the gap is
   allowed, while writes overlapping the partition, the whole disk and the
   partition itself are refused.

Open question

The area outside every partition includes the partition table (sector 0, and
the backup GPT at the end of the disk), so writing it is now allowed while
partitions are mounted. The mounted filesystems are not corrupted, since the
kernel keeps its in-memory table, but the next boot may see a different
layout. That is what full-disk images need, but if you would rather refuse
writes reaching the partition table, I can add that.

Testing

• cargo fmt --all -- --check, cargo clippy --locked --all-features --all --tests -- -D clippy::all
  and cargo test --locked --all --all-features pass.
• The root-only tests, run in a privileged container: the utils::fs ones
  pass, including the new one, and so do the copy installer ones. The
  flash, tarball, ubifs and mtd ones fail the same way on master in that
  environment (they need modprobe mtdram/nandsim).
• A mutation check: going back to treating every partition as reached makes
  the new loop device test fail.
• On hardware (am335x, eMMC, active/inactive tarball update with MLO and
  U-Boot as raw objects on /dev/mmcblk1), the first commit installs the
  package with the rootfs mounted.