fix(sft): enable max-length filtering for messages datasets

[none0663]

Summary

• Extend SFT long-sample filtering to support messages-style prompts (list[dict]) using chat-template tokenization.
• Prevent overlong samples from reaching dynamic batching and violating max_tokens_per_gpu token budget.

Why

Previously, max-length filtering mainly covered text-style prompts.
For messages datasets, overlong samples could pass through and later fail in training-time micro-batch partitioning (or cause OOM).

Test plan

• python3 -m py_compile slime/slime/utils/data.py
• Run SFT with --rollout-max-prompt-len and confirm log includes:
  Filtered X samples longer than max_length=...
• Verify no runtime failure from over-budget single samples in dynamic batching.