[fix][data] Apply chat-template kwargs during filtering

[bvolpato]

Summary

PromptDataset filtered examples without generator.chat_template_kwargs, while generation applied those options later. Settings such as enable_thinking can change rendered prompt length, so preprocessing could accept an over-length prompt or reject a valid one.

Changes

• Add optional chat_template_kwargs to PromptDataset.
• Pass generator chat-template options to training and evaluation datasets.
• Apply those options during prompt-length filtering.
• Share chat-template merging across filtering and non-batched generation, with explicit call controls taking precedence over configured kwargs.

Empty kwargs preserve existing behavior. Dataset records, collated batches, and inference contracts do not change.

Testing

• Full SkyRL Train CPU suite (1464 passed, 7 skipped, 5 deselected)
• Dataset suite (9 passed)
• Generator suite (32 passed)
• Collision regressions for thinking mode and explicit add_generation_prompt, return_dict, and tokenize controls (3 passed)
• pre-commit run --all-files
• git diff --check upstream/main

[SumanthRH]

@bvolpato Thanks! is this ready for review?

[bvolpato]

yes, ready for review now. I rebased it onto current main, reran the dataset suite and full pre-commit, and updated the PR description.

[gemini-code-assist]

Code Review

This pull request introduces support for passing custom chat_template_kwargs to the PromptDataset and its prompt filtering helper function, allowing configuration of the tokenizer's chat template application (e.g., enabling thinking tokens). Corresponding tests were added to verify this behavior. The review feedback suggests filtering out conflicting keys (add_generation_prompt, return_dict, tokenize) from chat_template_kwargs to prevent potential TypeErrors and logic breakage.