fix: wrap image bytes in BytesIO for gpt-image edit endpoint

The OpenAI SDK's legacy multipart path only accepts dall-e-2 when
raw bytes are passed. Wrapping in io.BytesIO with a name attribute
routes through the newer path that supports gpt-image-* models.

Also removes output_format from the edit call as that endpoint
does not support it.

tests/test_providers.py

@@ -7,6 +7,7 @@ Mock-heavy tests produce many Any-typed expressions from MagicMock.
 from __future__ import annotations
 
 import base64
+import io
 from pathlib import Path
 from unittest.mock import AsyncMock, MagicMock, patch
 
@@ -480,8 +481,10 @@ class TestOpenAIImageProvider:
             )
 
             call_args = mock_client.images.edit.call_args
-            # Single reference image should be passed as raw bytes
-            assert call_args.kwargs["image"] == b"reference data"
+            # gpt-image-* models pass a BytesIO with a name attribute
+            img_arg = call_args.kwargs["image"]
+            assert img_arg.read() == b"reference data"
+            assert hasattr(img_arg, "name")
 
         output = project_dir / "out.png"
         assert output.exists()
@@ -514,9 +517,9 @@ class TestOpenAIImageProvider:
             )
 
             call_args = mock_client.images.edit.call_args
-            # Multiple reference images should be passed as a list of bytes
-            image_arg: list[bytes] = call_args.kwargs["image"]
+            # gpt-image-* models pass a list of BytesIO with name attributes
+            image_arg: list[io.BytesIO] = call_args.kwargs["image"]
             assert isinstance(image_arg, list)
             assert len(image_arg) == 2
-            assert image_arg[0] == b"ref1 data"
-            assert image_arg[1] == b"ref2 data"
+            assert image_arg[0].read() == b"ref1 data"
+            assert image_arg[1].read() == b"ref2 data"