fix(executor): partition grouped global batches

[sylvesterkaczmarek]

Summary

• keep grouped global-batch operators hash-partitioned even when actor concurrency is 1
• retain at least the dataset's existing block count so unrelated source groups are not coalesced into one oversized Arrow block
• preserve single-block behavior for truly ungrouped global-batch operators
• add regression coverage for existing block counts, actor-pool concurrency, and ungrouped stages

Fixes #2643

Validation

• python -m compileall -q nemo_retriever/src/nemo_retriever/graph/executor.py nemo_retriever/tests/test_global_batch_partitioning.py
• git diff --check

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge with no actionable issues identified.

Summary

• This PR updates Ray Data global-batch partitioning so grouped operators remain hash-partitioned even with actor concurrency of one. It preserves at least the dataset’s existing block count for grouped stages, expands partitions when actor concurrency is higher, and retains single-block behavior for ungrouped stages. Unit tests cover these three partitioning cases.

_{Reviews (1) · Last reviewed commit: "fix(executor): partition grouped global ..."}