refactor(ideogram4): register pipeline and load via from_pretrained

Register the subclass with generic.set_pipeline and build it with
from_pretrained, passing the SDNQ transformers and shared text encoder
while the vae, scheduler, and tokenizer load from the repo.

modules/shared_items.py

@@ -40,7 +40,6 @@ pipelines = {
     'HiDream': getattr(diffusers, 'HiDreamImagePipeline', None),
     'HunyuanDiT': getattr(diffusers, 'HunyuanDiTPipeline', None),
     'HunyuanImage': getattr(diffusers, 'HunyuanImagePipeline', None),
-    'Ideogram4': getattr(diffusers, 'Ideogram4Pipeline', None),
     'JoyEdit': getattr(diffusers, 'JoyImageEditPipeline', None),
     'Kandinsky21': getattr(diffusers, 'KandinskyCombinedPipeline', None),
     'Kandinsky22': getattr(diffusers, 'KandinskyV22CombinedPipeline', None),
@@ -70,6 +69,7 @@ pipelines = {
     'FLEX': None,
     'HiDreamO1': None,
     'HunyuanImage3': None,
+    'Ideogram4': None,
     'Lens': None,
     'LuminaDiMOO': None,
     'Meissonic': None,