fix: add default writer_batch_size to prevent pyarrow offset overflow

[dieutx]

Summary

Closes #230.

• Datasets with large strings (e.g. serialized test cases in coding datasets like DeepCoder's lcbv5 subset) cause pyarrow offset overflow during dataset.map() calls because the default batch size produces Arrow arrays exceeding the 2 GB limit.
• Sets writer_batch_size=200 as a default in Environment.__init__ (propagated to all internal .map() calls via map_kwargs) and in the standalone format_dataset / load_example_dataset utilities.
• Users can still override writer_batch_size by passing it explicitly in map_kwargs.

Changes

• verifiers/envs/environment.py -- copy map_kwargs to avoid mutating the shared mutable default, then setdefault("writer_batch_size", 200).
• verifiers/utils/data_utils.py -- same default in format_dataset(), and explicit writer_batch_size=200 in load_example_dataset()'s .map() call.

Test plan

• Existing tests pass (uv run pytest tests/test_env_group.py tests/test_message_utils_multimodal.py)
• uv run ruff check and uv run ruff format clean

---

Note

Medium Risk
Changes default HuggingFace Dataset.map() batching across environment dataset formatting and example dataset utilities, which could affect performance/memory and any callers relying on previous implicit defaults.

Overview
Prevents pyarrow ArrowInvalid: offset overflow on datasets with very large string fields by defaulting HuggingFace .map() writer_batch_size to 200.

Environment now copies map_kwargs (avoiding mutation of the default dict) and applies the default via setdefault, while format_dataset() and load_example_dataset() apply the same default (including an explicit writer_batch_size=200 on the example preprocessing .map()). Documentation adds an FAQ explaining the error and how to override writer_batch_size when needed.

^{Written by Cursor Bugbot for commit e08dc65. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}