Enable access to reasoning_text and tool_calls in post-hoc LLM judges via flexeval_file.

[junya-takayama]

flexeval_lm passes LMOutput objects to Metric.evaluate(), giving metrics access to the full generation output including reasoning_text and tool_calls. However, flexeval_file, used for post-hoc evaluation of saved results, only passed the plain lm_output string, discarding all other fields. This made it impossible to apply LLM-judge metrics to reasoning content or tool call data after the fact.

This PR closes that gap: flexeval_file now reconstructs LMOutput from the saved fields, making post-hoc evaluation with LLM judges behave consistently with online evaluation via flexeval_lm.

Implementation summary

• evaluate_from_data() now reconstructs LMOutput objects from flat eval data dicts, picking up raw_lm_output, reasoning_text, finish_reason, tool_calls, and tool_call_validation_result fields
• LLM judge metrics (ChatLLMScore, ChatLLMGEvalScore, ChatLLMLabel) now receive LMOutput objects directly instead of plain strings, enabling Jinja2 templates to access fields like {{ lm_output.reasoning_text }}
• Added LMOutput.__str__ so that existing templates using {{ lm_output }} continue to render the text field without modification (backward compatibility)