feat: implement caching support for RNTuple reading (#1513)

* Implemented reading cache for RNTuples

* Fixed issue with nonzero entry_start for deferred columns

* Better test

* Reverted granularity of page ranges

* Perform all reading operations in place

* A bit of cleanup

* Move cache to the cluster level

* Fix cuda tests

* Copilot fixes

---------

Co-authored-by: Ianna Osborne <ianna.osborne@cern.ch>

Author: ariostas

src/uproot/behaviors/RNTuple.py

@@ -23,6 +23,7 @@
 import uproot.language.python
 import uproot.source.chunk
 from uproot._util import no_filter, unset
+from uproot.behaviors.TBranch import _regularize_array_cache
 
 
 def iterate(
@@ -615,7 +616,7 @@ def arrays(
         entry_start=None,
         entry_stop=None,
         decompression_executor=None,  # TODO: Not implemented yet
-        array_cache="inherit",  # TODO: Not implemented yet
+        array_cache="inherit",
         library="ak",  # TODO: Not implemented yet
         backend="cpu",
         interpreter="cpu",
@@ -659,7 +660,7 @@ def arrays(
                 is used. (Not implemented yet.)
             array_cache ("inherit", None, MutableMapping, or memory size): Cache of arrays;
                 if "inherit", use the file's cache; if None, do not use a cache;
-                if a memory size, create a new cache of this size. (Not implemented yet.)
+                if a memory size, create a new cache of this size.
             library (str or :doc:`uproot.interpretation.library.Library`): The library
                 that is used to represent arrays. Options are ``"np"`` for NumPy,
                 ``"ak"`` for Awkward Array, and ``"pd"`` for Pandas. (Not implemented yet.)
@@ -725,6 +726,8 @@ def arrays(
             [c.num_entries for c in clusters[start_cluster_idx:stop_cluster_idx]]
         )
 
+        array_cache = _regularize_array_cache(array_cache, self.ntuple._file)
+
         form, field_path = self.to_akform(
             filter_name=filter_name,
             filter_typename=filter_typename,
@@ -747,17 +750,22 @@ def arrays(
                 clusters_datas,
                 start_cluster_idx,
                 stop_cluster_idx,
-                pad_missing_element=True,
             )
 
         for key in target_cols:
             if "column" in key:
                 key_nr = int(key.split("-")[1])
+                # Find how many elements should be padded at the beginning
+                n_padding = self.ntuple.column_records[key_nr].first_element_index
+                n_padding -= cluster_starts[start_cluster_idx]
+                n_padding = max(n_padding, 0)
                 if interpreter == "cpu":
-                    content = self.ntuple.read_col_pages(
+                    content = self.ntuple.read_cluster_range(
                         key_nr,
-                        range(start_cluster_idx, stop_cluster_idx),
-                        pad_missing_element=True,
+                        start_cluster_idx,
+                        stop_cluster_idx,
+                        missing_element_padding=n_padding,
+                        array_cache=array_cache,
                     )
                 elif interpreter == "gpu" and backend == "cuda":
                     content = content_dict[key_nr]