resize: make interrupted resizes safe to resume

[eriknordmark]

Fixes #13.

Problem

A resize interrupted (e.g. the machine reboots mid-operation) and then re-run
could corrupt the disk. planResizes recomputes the plan from the current
on-disk state each run; once the relocated *_resized2 partitions have been
created, that space is occupied, the grows no longer fit, and a second
shrink of the shrink partition is planned — driving its target size negative.

Fix

• Resume-aware planResizes: a grow whose relocated *_resized2 partition
  already exists is reused in place and excluded from the space/shrink
  re-planning, so no second shrink is computed.
• copyFilesystems copy-skip: when the target already holds a filesystem
  matching the source (sync.CompareFS), the reformat+recopy is skipped;
  otherwise it reformats (over an empty or non-matching target) and recopies.
• Idempotent updatePartitions finalize: replaces the non-idempotent
  swapPartitions + removePartitions/removeAndRenumberPartitions finalize
  with a single partition-table write that relabels/reindexes each relocated
  target and removes the original, keyed on the stable on-disk start offset, so
  a re-run converges instead of undoing a completed operation. This makes the
  finalize step itself safe to resume.

Tests (resume_test.go)

TestRunResumeAfterInterruption interrupts the pipeline after each step and
re-runs to completion, asserting the final disk matches an uninterrupted run:
afterShrinkFilesystems, afterShrinkPartitions, afterCreatePartitions,
midCopyTargetFsCreated, midCopyTargetFsHasStaleFile, afterCopyFilesystems,
and afterUpdatePartitions (whole resize completed → re-run is a no-op) — in
both renumber and preserveNumbers modes. TestUpdatePartitions covers the
finalize transformation directly. TestRunAbortsOnFsckFailure asserts the
resize aborts on e2fsck failure rather than shrinking a broken filesystem.

swapPartitions/removePartitions/removeAndRenumberPartitions remain defined
but are no longer called by resize.

These cases run a real shrink/copy of a multi-GB fixture, so they are guarded by
testing.Short() and the CI test timeout is raised to 30m.

[eriknordmark]

@deitch ready for review. Fixes #13: an interrupted resize that was re-run could corrupt the disk (planResizes re-planned a second shrink once the relocated partitions existed, driving the shrink partition's target size negative). Makes planResizes resume-aware and adds a CompareFS-based copy-skip, plus restart-safety tests that interrupt the pipeline at each step and re-run to completion. Stacked on #6/#11 + the go-diskfs bump; the end-to-end tests need diskfs/go-diskfs#410, so CI here stays red until that merges and the pin updates (verified green locally against #410).

[deitch]

I like this. idempotency always was the eventual goal. It does need a rebase though.