fix: raise ContentFilterFinishReasonError on refusal responses

[danylokv-lang]

Summary

Fixes parsing error where refusal responses caused a confusing pydantic_core.ValidationError instead of a proper SDK exception.

Problem

When content moderation is triggered, the model returns a plain-text refusal like "I'm sorry, but I cannot assist with that request." instead of valid JSON. This refusal text was passed directly to Pydantic validation:

for item in output.content:
    if item.type != "output_text":
        content_list.append(item)
        continue
    # Refusal text reaches here and fails JSON parsing!
    parsed = parse_text(item.text, text_format=text_format)

Since refusal text is not valid JSON, users received a low-level pydantic_core.ValidationError with no indication that the issue was content moderation.

Solution

Added a guard to detect refusal content before attempting Pydantic validation:

for item in output.content:
    if item.type == "refusal":
        from ..._exceptions import ContentFilterFinishReasonError
        raise ContentFilterFinishReasonError() 
    if item.type != "output_text":
        content_list.append(item)
        continue

This is consistent with how refusals are handled in _completions.py (line 102).

Changes

_responses.py: Added refusal detection that raises ContentFilterFinishReasonError

Test Plan

✅ Static analysis passes (no errors)
✅ Logic now surfaces moderation-triggered responses as ContentFilterFinishReasonError
✅ Consistent with chat completions refusal handling
Fixes #2834