Consolidate fl-tutorials dataset tooling into a shared datasets/ tree

[atriaybagur]

What

Consolidates the tutorial datasets' download/derive/enrich tooling into a shared fl-tutorials/datasets/ tree, with all outputs in one shared gitignored fl-tutorials/data/ root. Closes #1069.

Datasets are backend-agnostic (a tutorial's NVFLARE and Flower twins train on the same data) and many-to-one (spleen alone backs segmentation, evaluation and diffusion), but the tooling was scattered per backend and drifting: near-identical xray recipes in both backend Makefiles (different output CSV names, separate .hf caches — testing both backends downloaded the same data twice), and the spleen scripts living under one tutorial's utils/ with the Flower Makefile delegating cross-tree behind an apologetic comment.

How

• fl-tutorials/datasets/Makefile owns every dataset target (single copy): download-xray-data, download-spleen-data (FL_BACKEND-dispatched to the explicit download-spleen-msd-data / download-spleen-flip-format-data variants), download-arkplus-{finetuning,eval}-data, upload-spleen-labels. The fl-tutorials root Makefile forwards these there, so every documented entrypoint is unchanged (make -C fl-tutorials download-spleen-data NUM_CASES=41 [FL_BACKEND=flower], etc.).
• fl-tutorials/datasets/spleen/ takes the three spleen scripts and their uv project (moved from nvflare/image_segmentation/3d_spleen_segmentation/{utils/,pyproject.toml}; the project env existed solely for these utilities), plus a new download_spleen_flip_format_dataset.py for the FLIP-format variant. datasets/README.md documents each dataset's source, layout and consumers.
• Every download is a script, not an inline python -c: the xray, arkplus and FLIP-format spleen fetches each live in their dataset folder (xrays_mini_300/, arkplus/, spleen/) and run via uv run --no-project --with huggingface_hub; the Makefile targets just invoke them.
• One shared data root: both backends now read fl-tutorials/data/; one download serves both. The xray CSV gets one canonical name (dataframe.csv); the two spleen variants (MSD build: images/ + dataframe.csv; FLIP-format: accession-resources/ + sample_get_dataframe_response.csv + checkpoint) coexist under data/spleen/ — the FLIP-format download now removes only its own outputs instead of rm -rf data/spleen.
• Consumers re-pointed: the 7 nvflare .env.app files, flower/run-tutorial.sh (which now also fail-fasts on a missing dataframe file, not just the images dir), flip-api's e2e_smoke_spleen* enrichment wiring (SPLEEN_ENRICHMENT_CWD/CMD/LABELS_DIR), root CLAUDE.md/AGENTS.md, tutorial READMEs, fl-services/flower/README.md, the demo-video docs and the arkplus user guide. The upload script's remediation text now names the two variants instead of the two backends.
• Removes the accidentally committed self-referencing fl-tutorials/nvflare/data symlink.
• Strips the leftover Flower example-gallery frontmatter (tags/dataset/framework) from the two Flower spleen tutorial READMEs — it is metadata for Flower's examples website and renders on GitHub as a stray horizontal rule plus literal text.
• A side-effect fix: the create-flip-app-from-flower.rst local-testing example (../../data/spleen/... from the tutorial root) pointed at a path that didn't exist under the old layout — it is exactly right under the new one.

Verification

• make -C fl-tutorials download-xray-data and make -C fl-tutorials download-spleen-data FL_BACKEND=flower run end-to-end into the shared root, with the expected layouts, and git status stays clean afterwards (gitignore covers the new root).
• Target dispatch, NUM_CASES forwarding, and the per-variant SPLEEN_LABELS_DIR defaults verified via make -n command expansion; bash -n on run-tutorial.sh; list-tutorials works on both backends.
• make -C fl-tutorials test green (ruff + 112 passed, 12 expected skips — the count grew when develop was merged in: its new FLIP#1060 spleen case-selection tests were re-pointed at the moved datasets/spleen/ downloader as part of the merge).
• scripts/check_tutorial_sync.sh green; a repo-wide grep finds zero references to the old paths.
• Not run (large downloads, unchanged recipes beyond path prefixes): the MSD spleen and arkplus downloads; a live upload-spleen-labels against XNAT.

Sequencing

Merge before #1068 — the EHR risk-prediction tutorial then rebases onto this and moves its build_synthea_dataframe.py into datasets/synthea/, instead of adding a second cross-tree delegation.

Acceptance Criteria

Imported from issue #1069

• fl-tutorials/datasets/<name>/ owns each dataset's scripts + Make targets; no dataset scripts remain under per-tutorial utils/ and no duplicated download recipes remain in the per-backend Makefiles.
• Existing entry points keep working unchanged: make -C fl-tutorials download-xray-data|download-spleen-data|download-arkplus-*|upload-spleen-labels [FL_BACKEND=flower].
• One shared gitignored data root — downloading a dataset once serves both backends' tutorials (per-backend variants, e.g. the Flower pre-built FLIP-format spleen tree vs the NVFLARE MSD build, remain available as distinct targets/outputs).
• All consumers re-pointed: per-tutorial .env.app paths, flower/run-tutorial.sh mappings, flip-api e2e_smoke_spleen* defaults, root CLAUDE.md/AGENTS.md, READMEs, docs.
• make -C fl-tutorials test green; tutorial-sync CI green; downloads verified to land in the layout the tutorials expect.

[github-actions]

✅ Acceptance criteria have been automatically imported from the linked issue(s) and added to the PR description.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[garciadias]

Clean, mechanical, thoroughly self-verified refactor consolidating dataset tooling; all path rewrites across .env.app files, Makefiles, run-tutorial.sh, flip-api/Makefile, and the CLAUDE.md/AGENTS.md twin-file pair check out, no stale references remain, CI is green. One optional non-blocking nit left inline. Approving.