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

[bvolpato]

Problem

PromptDataset filters examples by calling tokenizer.apply_chat_template without generator.chat_template_kwargs. Generation later applies those kwargs. Options such as enable_thinking can change rendered prompt length, so filtering measures a different prompt from the one sent to inference. Over-length prompts may pass preprocessing and fail only at runtime; valid prompts can also be filtered using the wrong template mode.

Change

• add optional chat_template_kwargs to PromptDataset
• pass generator.chat_template_kwargs to both training and evaluation datasets
• use those kwargs in length-filter tokenization

Why this belongs in SkyRL

SkyRL owns preprocessing and passes these tokenizer options during generation. Length filtering must use same configured kwargs to enforce trainer.max_prompt_length against prompt that will run. Fix applies to any tokenizer kwarg that changes rendering, not one model family.

Compatibility

• default empty kwargs preserve existing rendering exactly
• custom chat-template selection is outside this change
• dataset records and collated batch format do not change
• no inference request or response contract changes
• configured kwargs are forwarded unchanged to tokenizer

Validation

• added focused regression coverage where thinking mode changes prompt from one token to four
• confirmed filter uses configured mode and rejects prompt at correct limit
• focused dataset tests: 2 passed
• changed-file Ruff, Black, and hardcoded-secret checks pass

CI note

Current full-tree Black failure is already present on main in tests/tinker/test_engine.py and is isolated in #2009.