fix: preserve standalone CLIP-Vision tensors for IP-Adapter (#1977)

[Maphist0]

Summary

Fix loading of standalone CLIP-Vision tensors used by SD1.x IP-Adapter models. Related to issue 1977

The global vision_model. unused-tensor filter was applied before component prefixes were added, causing raw vision_model.* tensors in standalone CLIP-Vision files to be skipped.

The filter now preserves those tensors only when loading with the clip_vision. prefix.

Related Issue / Discussion

issue 1977

Additional Information

SD1.5 + IP-Adapter:

• CLIP-Vision loaded successfully.
• IP-Adapter: 4 image tokens reported.
• Image generation completed successfully.

Log:

$ sd-cli -M img_gen \
  -m .../v1-5-pruned-emaonly.safetensors \
  --clip_vision .../clip-vision-model.safetensors \
  --ip-adapter .../ip-adapter_sd15.safetensors \
  --ip-adapter-image .../style.png \
  -p "a photo" --steps 4 -W 512 -H 512 --seed 42

loading clip_vision from '.../clip-vision-model.safetensors'
load .../clip-vision-model.safetensors using safetensors format
load .../ip-adapter_sd15.safetensors using safetensors format
Version: SD 1.x
IP-Adapter: 4 image tokens, strength 1.00
generate_image completed in 10.56s
save result image 0 ... (success)
1/1 images saved

SenseNova U1.5:

• U1.5 FLOW mode initialized successfully.
• Image generation completed successfully.
• 1/1 images saved.

Log:

$ sd-cli --model .../SenseNova-U1.5-8B-MoT-... \
  --prompt "a red cube on a white background" \
  --width 512 --height 512 --steps 20 --cfg-scale 4 \
  --flow-shift 3 --seed 42 --sampling-method euler --rng cuda

Version: SenseNova U1.5
running in SenseNova U1.5 FLOW mode
generate_image 512x512
sampling completed, taking 27.88s
generate_image completed in 27.90s
save result image 0 ... (success)
1/1 images saved

Checklist

• I have read and confirmed this PR follows the contribution guidelines.

[leejet]

Superseded by #1983.