Fix create-recore-config to actually provision the eMMC config

[eliasbakken]

Summary

Three bugs, all reported in #85:

1. Formatted a nonexistent partition node - mkfs.ext4 targeted /dev/mmcblk2boot0p1, but the kernel doesn't create child partition nodes for eMMC boot partitions. mount-config and every other reader loop-mount the raw device directly at offset 17408 - the filesystem is now created the same way, via a loop device at that offset, not via fdisk/a partition node that never existed.
2. metadata_csum (default-on in e2fsprogs 1.47) left no room for the dir-leaf checksum once the filesystem picked a 1K block size at this tiny size, so reads failed with Bad message. Disabled via -O ^metadata_csum,^64bit.
3. Wrote through a read-only mount - the create path used mount-config, which always mounts read-only (it's meant for readers). Now mounts the loop device directly, read-write, for the create path only.

Also fixed a related bug in get-recore-serial-number: it called plain umount instead of unmount-config to tear down the transient mount mount-config sets up via systemd-mount. Plain umount unmounts the filesystem but doesn't deregister the transient systemd unit, so a later mount-config call can fail with Unit mnt-config.mount was already loaded.

Testing

• make test-bats - all 13 pass (updated the existing get-recore-serial-number test to stub unmount-config instead of umount, matching the fix).
• go test ./... - no regressions.
• Live-tested on real hardware: extracted and backed up the board's existing (already-provisioned) config partition first, ran the fixed create-recore-config against the same serial number, and confirmed get-recore-revision/get-recore-serial-number both work and the downloaded calibration.json is byte-identical to the original.

Closes #85