fix: handle empty OpenAI-compatible text responses

[andyluo7]

Summary

Harden the OpenAI-compatible evaluation path so None or structured message.content values do not break lmms-eval postprocessing.

Changes

• normalize OpenAI-compatible message.content so None becomes an empty string and list-style structured content is flattened into plain text
• guard unwrap_generation_output() against raw None and GenerationResult(text=None)
• guard ConfigurableTask.process_results() so generate-until results are coerced to strings before .strip()

Why

Some OpenAI-compatible backends can legally return message.content = None or structured content blocks. Previously those values propagated into task postprocessing and caused crashes like:

AttributeError: 'NoneType' object has no attribute 'strip'

Validation

• verified the patched files parse cleanly
• confirmed OCRBench evaluation completes successfully after this fix