perf(data): parallelize NeMo Gym image encoding

[youngeunkwon0405]

What does this PR do?

Parallelizes local-image conversion in encode_images_in_examples without changing its public API or supported payload shapes.

The function now:

• groups image parts by their exact local source string;
• encodes each unique source once with a bounded ThreadPoolExecutor(max_workers=8);
• closes each decoded PIL image as soon as encoding finishes to release its pixel buffer;
• mutates the NeMo Gym payload only on the caller thread after all unique sources have encoded successfully; and
• reuses one data-URL string for every part sharing a source, reducing the payload serialized across the Ray boundary.

The worker count is capped at eight to bound peak decoded-image memory. In a CPU microbenchmark using 1024 x 1024 PNGs, the thread pool provided approximately 3.7-3.9x speedup. Deduplication is a separate gain: GRPO repeats each prompt per generation, so a group size of N previously encoded the same image approximately N times and now encodes it once.

HTTP(S) and existing data: URLs remain pass-through values, and local paths plus file:// URLs keep the existing PNG data-URL behavior.

Issues

None.

Validation

• uvx --from ruff==0.9.9 ruff check nemo_rl/data/multimodal_utils.py tests/unit/data/test_multimodal_image_encoding.py
• uvx --from ruff==0.9.9 ruff format --check nemo_rl/data/multimodal_utils.py tests/unit/data/test_multimodal_image_encoding.py
• uv run --extra nemo_gym --group test pytest -q --noconftest tests/unit/data/test_multimodal_image_encoding.py - 7 passed on the initial patch
• Refreshed GitHub Lint check on 634f5450fecf30ba09c6495261586e01ec370ee2 - passed, including Pyrefly

Before your PR is "Ready for review"

• Read and followed the contributor guidelines
• Added focused tests
• Ran focused Linux unit tests
• No documentation update is needed because this is an internal performance change with no new configuration or public API

[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.

[youngeunkwon0405]

/ok to test 6957387

[youngeunkwon0405]

/ok to test 634f545

[aroshanghias-nvd]

One non-blocking operational concern:

[youngeunkwon0405]

/ok to test 462ff92

[DanialTaheri]

looks good to me

[DanialTaheri]

@yfw @terrykong can we merge this? It is a harmless optimization for image encoding.

[youngeunkwon0405]

Hi @terrykong, can I get your help with merging this PR? It is a ~20-line logic change that parallelizes image encoding for the VLM models. I already did it through review-pr-team, the CI run passed, and I got approval from @DanialTaheri. Thanks for your help always!