Add ReturnFieldsRetriever: concurrent scatter/gather utility for per-hit return fields

[zihanx]

Description

Adds a reusable utility for fetching "return fields" — one value per hit for a set of
global doc IDs — concurrently across leaves, returning results 1:1 in the caller's input
order. This is the follow-up to the ReaderUtil.partitionByLeaf scatter/gather API
#16496; it builds the retriever the ordinal tracking was designed for.

Motivation

Fetching field values for a page of hits (in ranking order) is a common need, and getting
the leaf partitioning + concurrency + input-order reassembly right is tricky and easy to
get subtly wrong. Today callers hand-roll this. This packages it once.

What's included

Two classes:

• ReturnFieldsRetriever — the generic engine. Owns only the reusable machinery:
  scatter doc IDs to leaves via ReaderUtil.partitionByLeaf, run each leaf as an
  independent task on a supplied Executor (via TaskExecutor), and gather per-leaf
  results back into input order. It is agnostic about what is produced per hit: the
  caller supplies a LeafVisitorFactory<T> that turns a document into a value of type T.

  public static <T> T[] retrieve(
      IndexReader reader, int[] globalDocIds,
      LeafVisitorFactory<T> factory, IntFunction<T[]> arrayFactory, Executor executor)

• DoubleValuesRetriever — a thin DoubleValuesSource-specific specialization built
  on the engine. Adapts each source to a LeafVisitor<double[]>:

  public static double[][] retrieve(
      IndexReader reader, int[] globalDocIds, DoubleValuesSource[] sources, Executor executor)

Design notes

• Concurrency contract: leaves run concurrently, so newLeafVisitor is called once per
  leaf and must return a fresh, single-threaded visitor. Each leaf writes disjoint slots of the result array,
  so no synchronization is needed.
• Input order preserved: result[i] is the value for globalDocIds[i]. The input
  array is not mutated.