feat(image): adopt Docker v29 output for images list

[ekalinin]

Make the default nerdctl images output match Docker v29: a collapsed view with
IMAGE, ID, DISK USAGE, CONTENT SIZE and EXTRA columns, an "In Use" (U)
indicator, and <untagged> for dangling images. Multi-platform images are collapsed
into a single row with aggregated disk and content size.

Details that follow Docker's behavior:

• DISK USAGE is the content store blobs plus the unpacked snapshots (Docker's
  Total), CONTENT SIZE is the blobs alone. Both are formatted with 3 significant
  digits, like Docker.
• In use is resolved by image target digest, the way Docker matches containers to
  images, so every reference to a used target is flagged, not only the one the
  container was created from.
• The "In Use" legend is only printed when the output is a terminal, so piped and
  redirected output stays clean.
• Rows are ordered by image reference with untagged images last, rather than by
  creation time.
• The ID column is the OCI digest of the image target, which matches Docker's ID
  with the containerd image store (it differs only from the legacy graphdriver
  image ID). The help text and docs are updated accordingly.

The new view is used only for the bare command. Passing --format, --quiet,
--no-trunc, --digests or --names falls back to the legacy table (REPOSITORY,
TAG, IMAGE ID, CREATED, PLATFORM, SIZE, BLOB SIZE), so existing scripts and
templates keep working, including their creation-time ordering. This mirrors Docker's
own shouldUseTree fallback.

Since the default output now matches Docker, the images tests also run against the
Docker target; only the nerdctl-specific --names subtest stays gated. Tests that
assert on the default images output for untagged images (image prune/remove and
build-without-tag) are updated to expect <untagged>.

The expanded per-platform --tree view is left for a follow-up.

Closes #5027

[AkihiroSuda]

Thanks

[ekalinin]

Force-pushed. Both threads above are addressed (DISK USAGE is now Total = content + unpacked, and the ID note is updated). On top of that, two more Docker-compat fixes that I found while re-reading the diff:

1. "In Use" is now resolved by image target digest instead of image name

moby matches containers to images by digest (daemon/containerd/image_list.go#L338):

  if c.ImageManifest != nil && c.ImageManifest.Digest == target.Digest {

nerdctl was keying the in-use set on container.Image().Name(), so only the reference the container had been created from got the U flag. With nerdctl tag alpine:latest myalias:v1 and a container running off alpine:latest, the myalias:v1 row stayed empty even though it points at the same target. Now every reference to a used target is flagged. The "In use" integration test tags the image under a second name and asserts that all rows pointing at that target are marked, so it also covers the Docker target (where those names are collapsed into a single row).

2. Row ordering in the new view now follows docker/cli

docker/cli's tree view sorts by image name lexicographically and places untagged images last, while nerdctl kept its creation-time ordering. The new view is now sorted by the rendered IMAGE value with <untagged> last. This is applied to the new view only: the legacy table keeps the creation-time ordering that --format consumers (and the CheckCreatedTime subtest) rely on. Added a unit test for the comparator.

---

One known divergence left, which I'd rather fix separately - let me know if you want it here

CONTENT SIZE sums the descriptor sizes declared in the manifest without checking that the blobs are actually present locally, and skips attestation manifests entirely. moby walks only the present content (walkPresentChildren, ImageManifest.PresentContentSize) and adds it to TotalSize before the pseudo-image / attestation check, so attestation content is counted.

So nerdctl can over-report for partially pulled images - which is not exotic here, given the lazy-pulling snapshotters (stargz, nydus, SOCI, overlaybd) - and under-report when attestations are present. This math predates this PR (it is the same code behind today's BLOB SIZE column), but the new view now surfaces it in two columns instead of one, so it is worth calling out. I'll open a separate issue unless you prefer it folded in.

PTAL.