[Jeronymous]

Report how often a model's generation is cut off by reaching a token budget, distinguishing the answer from a thinking model's reasoning.

• ModelResponse gains per-generation output_truncated (answer / whole generation) and reasoning_truncated (two-phase reasoning phase) flags.
• vLLM and SGLang set them from the native finish reason; the two-phase thinking path sets them from the per-phase budgets, so they work for reasoning models on any backend.
• New Metrics.generation_truncated and Metrics.thinking_truncation report a rate (mean) and a count (sum).
• Both are added automatically to every generative task via LightevalTaskConfig, so no task needs editing by hand.

Not covered (reported as 0): non-thinking transformers/sglang single-phase stop-sequence stops have no per-sample finish reason, and remote endpoints.