fix: restore model_hijack.clip after LDSR upscale

LDSR calls sd_hijack.model_hijack.hijack() on its own diffusion model,
which overwrites model_hijack.clip with the LDSR cond_stage_model (a
VQ-VAE autoencoder). After upscaling completes, clip remains set to
this wrong object, causing token counters to fail with:
  AttributeError: 'Identity' object has no attribute 'process_texts'

Save and restore model_hijack.clip around the upscale operation so the
original SD model's CLIP encoder is preserved.

Fixes #12942

extensions-builtin/LDSR/ldsr_model_arch.py

@@ -104,6 +104,7 @@ class LDSR:
         return logs
 
     def super_resolution(self, image, steps=100, target_scale=2, half_attention=False):
+        previous_clip = sd_hijack.model_hijack.clip
         model = self.load_model_from_config(half_attention)
 
         # Run settings
@@ -151,6 +152,8 @@ class LDSR:
         gc.collect()
         devices.torch_gc()
 
+        sd_hijack.model_hijack.clip = previous_clip
+
         return a