docs: correct reverse KL definition in OPD guide

[XiXiphus]

Summary

• correct the reverse-KL direction to D_KL(student || teacher) and make the student-sampled expectation explicit, following the Thinking Machines Lab definition
• distinguish the full token-distribution KL from slime’s per-sampled-token Monte Carlo contribution and show how it modifies the base advantage
• clarify pure-distillation behavior and the tokenizer/vocabulary compatibility requirement for an external SGLang teacher
• keep the English and Chinese guides aligned

Validation

• bash docs/build.sh en (succeeded; only pre-existing warnings elsewhere in the docs)
• bash docs/build.sh zh (succeeded; only pre-existing warnings elsewhere in the docs)
• uvx pre-commit run --files docs/en/advanced/on-policy-distillation.md docs/zh/advanced/on-policy-distillation.md