perf: build the RNTuple form once and prune it per field

[kadircanyildirm-crypto]

Part of #1637.

to_akform rebuilt the form from the field records on every call, so reading a wide RNTuple one column at a time — which is what coffea does — rebuilt the whole form once per column. For the 16-event, 1491-field NanoAOD skim in the issue that is 1579 builds.

This builds the unfiltered form once per RNTuple, keeps it there, and serves later requests by pruning it to the requested keys. A field takes the same cached form, walks down to its own path and prunes from there, so the per-column path comes off the cache as well — that is the path coffea actually uses, since it asks fields rather than the ntuple. For a field under a collection the form is still built around the outermost collection on its path and field_path is returned as before; ak_add_doc still takes the original path.

Measured on main (c979803), Windows, Python 3.12, awkward 2.13.0, coffea 2026.7.0, reading the skim back through NanoEventsFactory:

	before	after
read-back	252.0 s	32.1 s / 36.2 s
full form builds	1579	1

The same data written as a TTree reads in 3.8–4.5 s, so this does not close the gap to TTree by itself. What is left is a different effect, measured and described in #1637 (comment).

Checks:

• all 1491 leaf columns are bit-identical to the ones built the old way (ak.to_buffers, compared byte for byte)
• pytest -k "rntuple and not dask": 146 passed, 0 failed — 143 on main, plus the 3 added here
• full suite identical to main: 1032 passed, 29 failed, 117 skipped either way, the 29 all in the dask test files because distributed is missing in my environment
• the added test asserts that 200 field requests cause one form build, and that the pruned form equals the rebuilt one for filters, subfields and ak_add_doc

Disclosure, per CONTRIBUTING: this was written with Claude Code. The measurements above are from runs on my own machine against this branch.

🤖 Generated with Claude Code

https://claude.ai/code/session_01B6b4SK8hGvDHGAx1W6KqGJ

[codecov]

Codecov Report

❌ Patch coverage is 80.89888% with 17 lines in your changes missing coverage. Please review.
✅ Project coverage is 75.38%. Comparing base (c979803) to head (8bf0c57).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/uproot/behaviors/RNTuple.py	80.89%	10 Missing and 7 partials ⚠️

❌ Your patch check has failed because the patch coverage (80.89%) is below the target coverage (98.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

Files with missing lines	Coverage Δ
src/uproot/behaviors/RNTuple.py	74.69% <80.89%> (-0.09%)	⬇️