[Perf/Fix] Reimplement Batched CFG Forward for Bagel (#4098)

Signed-off-by: Alex Brooks <albrooks@redhat.com>
Co-authored-by: 汪志鹏 <wangzhipeng628@gmail.com>

tests/distributed/omni_connectors/test_bagel_shared_memory_connector.py

@@ -7,7 +7,7 @@ End-to-end tests for Bagel with shared memory connector: img2img and text2img.
 - img2img: validates output vs reference pixels within a ±10 tolerance.
 - text2img: validates output vs reference pixels within a ±5 tolerance
   (equivalent to `examples/offline_inference/bagel/end2end.py` with
-  text2img modality and 15 steps).
+  text2img modality and 14 steps).
 """
 
 import os
@@ -46,7 +46,6 @@ REFERENCE_PIXELS = [
     {"position": (256, 256), "rgb": (181, 201, 221)},
 ]
 
-
 # text2img reference pixels (aligned with offline `bagel/end2end.py` text2img, 15 steps)
 # "Generated with seed=52, num_inference_steps=15,
 # prompt='A cute cat'"