BUG: GDCMImageIO/ImageSeriesReader collapse multi-segment DICOM SEG objects to a single 3D volume

## Summary

`GDCMImageIO` / `ImageSeriesReader` read a DICOM Segmentation Storage (SEG) object as a single 3D volume and have no concept of per-segment sub-stacks. When a SEG file contains **multiple segments** — which may have **different per-segment z-extents (frame counts)** — ITK collapses them into one 3D output, silently losing all but one segment's geometry/labeling.

This is a **structural limitation**, distinct from the single-bit decode defect tracked in #6641 (which is about pixel *values* on a single volume). It was raised by **@issakomi** on PR #6639.

## Steps to reproduce

1. Generate a multi-segment SEG (e.g. TotalSegmentator output, which uses [highdicom](https://github.com/ImagingDataCommons/highdicom)) containing ≥2 segments with differing z-extents.
2. Read it with `itk::ImageSeriesReader` + `GDCMImageIO`.

**Expected:** each segment recovered as its own labeled volume (or an API that exposes per-segment sub-stacks), as pydicom/highdicom-based readers do.

**Actual:** a single 3D volume; the additional segment(s) are dropped.

## Context

Raised in the PR #6639 review thread — https://github.com/InsightSoftwareConsortium/ITK/pull/6639#issuecomment-4976646767 — where @issakomi noted they can read all segments correctly via highdicom, while ITK IO outputs just one 3D file. Confirmed by the PR author as a real, separate, larger-scope limitation than the single-bit decode fix in that PR.

## Related

- #6641 — single-bit SEG mis-decode (0–255 bytes vs binary 0/1); different bug, same file family
- #6642 — highdicom cross-validation testbed proposal
- PR #6639 — single-bit contiguous decode + PixelSpacing multiplicity fix

Start with GDCMImageIO and ImageSeriesReader, then reproduce the issue using a multi-segment DICOM SEG with differing z-extents, such as TotalSegmentator output. Done means each segment is recovered as its own labeled volume, or the API exposes the per-segment sub-stacks without dropping segments.

Written by the indexing model from the issue text.