feat: Pydantic RNTuple extension and Iterative/FuturesBackends

[NJManganelli]

Adds RNTuple capability to preprocessing, plus alternative backends (besides dask) to permit different processing modes. The backend design is intended to potentially interface with the coffea.compute refactor in the future.

This has been largely implemented by Claude Fable 5 + Opus 4.8, with hands-on guidance from me

This PR is rebased to #1528

Todos:

• cleanup filespec.ipynb to better demonstrate the method-chaining for slicing etc
• decide on whether this backends approach makes sense right now
• full human review of code

[NJManganelli]

🤖 AI text below 🤖

Follow-up: exercising the RNTuple preprocessing + switchable backends with the bundled test files.

This PR adds RNTuple support (auto-detected, plus a strict preprocess_rntuple) and dask-free iterative/futures backends. Here's a minimal way for a maintainer to exercise both against the repo's own samples, going through the top-level DataGroupSpec API (no hand-built filespecs):

from coffea.dataset_tools import DataGroupSpec, preprocess, preprocess_rntuple

# One group mixing a TTree dataset and an RNTuple dataset; preprocess() auto-detects
# TTree vs RNTuple per file. backend="iterative" runs dask-free (single process).
fileset = DataGroupSpec({
    "ttree_dy":   {"files": {"tests/samples/nano_dy.root": "Events"}},
    "rntuple_dy": {"files": {"tests/samples/nano_dy_rntuple.root": "Events"}},
})
available, _ = preprocess(fileset, step_size=15, backend="iterative")
for name, ds in available.items():
    fs = next(iter(ds.files.values()))
    print(name, fs.steps, fs.num_entries)
# ttree_dy   [[0, 14], [14, 28], [28, 40]] 40
# rntuple_dy [[0, 14], [14, 28], [28, 40]] 40   <- identical steps/form via RNTuple path

# preprocess_rntuple asserts every object is an RNTuple (alias for
# preprocess_root(..., require_rntuple=True)); a TTree in the group would raise
# a clear ValueError (subject to skip_bad_files/file_exceptions). Here on a thread pool:
rnt_only = DataGroupSpec({"rn": {"files": {"tests/samples/nano_dimuon_rntuple.root": "Events"}}})
avail_rnt, _ = preprocess_rntuple(rnt_only, backend="futures")
print(list(avail_rnt["rn"].files))

The iterative and futures backends produce a DataGroupSpec byte-identical (decoded-form-wise) to the dask backend, including the saved union form, for both TTree and RNTuple — so this runs with or without the dask stack installed. tests/test_dataset_tools_backends.py pins that equivalence (and tests/test_dataset_tools.py::test_preprocess_rntuple pins the RNTuple form against uproot.dask).

Three follow-on features build on this branch as stacked draft PRs against master (each includes this branch's commits until it merges, then collapses to just its own feature): union forms via DatasetSpec addition #1599 (#1478), per-file metadata extraction #1600 (#1401), and a mutable-steps prototype #1601 (#1483).

---

Developed with Claude Fable 5.