feat: add agentic evaluator

[Yunnglin]

Summary

• add a single-use EvalScope Native evaluator for Twinkle Agentic protocol APIs and structural samplers
• preserve text, reasoning, tool calls, stops, multiple choices, and supported generation settings at the model boundary
• add sampler micro-batching, DP tail padding, and HTTP sampler SamplingParams serialization
• document installation, API/sampler usage, capability boundaries, and non-goals in Chinese and English

Validation

• python -m pytest tests/twinkle_agentic/evaluator tests/twinkle_agentic tests/twinkle_client/test_client_multi_turn_rollout.py -q
  • 208 passed, 4 skipped
• real offline EvalScope Native general_qa evaluation passes with both a fake protocol API and fake sampler
• python -m pip wheel --no-deps .
• git diff --check

Scope

No GPU model or external endpoint is required for the included validation. EvalScope benchmark-specific extras and real sampler/API smoke tests remain caller-provided runtime configuration.