fix: keep the object cache in the graphed TTree source

[lgray]

🤖 AI text below 🤖

Fixes #1730

_GraphedTTreeSource._open_directory no longer passes object_cache=None, so ReadOnlyFile's default (object_cache=100, as in uproot.open) applies. The file is already opened once per worker through resources.open_once; with its object cache, later partitions of that file reuse the TTree instead of deserializing it again. array_cache=None stays, so no arrays are held across partitions.

No caller setting is overridden: an explicit object_cache= passed to uproot.graphed is already rejected at construction by _util.regularize_object_path, which uproot.dask shares.

Timing

Events tree of Run2012B_SingleMu.root (CMS Open Data NanoAOD, 16 GB, 86 branches), first 2^21 entries as 4 explicit steps of 2^19; gak.sum(MET_pt) through graphed.aggregate_plan + SequentialRunner, median of 5 runs after one warm-up; Apple M1 Max, Python 3.13, graphed 0.0.5.

	main (ca3a8a2)	this PR
TTree reads for 4 partitions	4	1
open_tree per partition (ms)	20.9, 20.7, 19.9, 21.2	21.0, 0.0, 0.0, 0.0
sum of MET_pt, 4 partitions (s)	0.121	0.057

Test

tests/test_1730_graphed_tree_cache.py: two partitions of one file, opened through one LocalResources, get the same TTree object. It fails on main (assert second is first: two distinct TTree objects) and passes here.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 76.15%. Comparing base (ca3a8a2) to head (ededf64).

Additional details and impacted files

Files with missing lines	Coverage Δ
src/uproot/_graphed.py	100.00% <ø> (ø)

[lgray]

@ariostas Here's another quick fix from doing higher level tests with the nanoevents integration built in scikit-hep/coffea#1637.

[ariostas]

Seems reasonable to me. The size of a tree object increases with the number of branches, but even with a "standard" NanoAOD with O(1000) branches it's only like 10MB, so it's probably fine.

Thank you, @lgray!