docs: surface agent workflows and fix vLLM response parsing

[aoshen02]

Summary

• surface Vime's existing agentic rollout, sandbox, and coding-agent workflows in English and Chinese docs
• centralize aligned vLLM token-id/log-prob parsing
• fix stale SGLang response assumptions in the existing mem_agent and multi_agent examples
• intentionally keep Slime's examples/retool and examples/search-r1 out of Vime for now

Full cutoff audit

Compared the complete Slime tree at 41014d1f against Vime base f1d3c6b81a, and audited all 243 files touched since the previous Slime cutoff—not only multi-turn or sandbox files.

• explicit SGLang-to-vLLM counterparts were checked for engine, rollout, config, tests, and docs
• all remaining absent paths were classified as signed/explicit exclusions
• retool (8 files) and search-r1 (9 files) exist in Slime at the cutoff, but are deliberately excluded from this PR per review decision
• no additional unported shared runtime change was found

tools/convert_to_hf.py still references the removed upstream update_weight_utils, but the same stale import exists verbatim in Slime 41014d1f; it is not a Vime sync omission and remains outside this focused translation.

Validation

• pre-commit passed for all changed files
• PYTHONPATH=$PWD pytest -q tests/test_vllm_rollout.py (36 passed)
• git diff --check