add new SOTA Method: isotropic merging

[tanganke]

Example Usage:

Merge CLIP-ViT-B/32 models on eight image classification tasks using ISO-C, with a scaling factor of 1.5:

fusion_bench \
    fabric.loggers.name=iso_c \
    method=isotropic_merging/iso_c \
    method.scaling_factor=1.5 \
    modelpool=CLIPVisionModelPool/clip-vit-base-patch32_TA8 \
    taskpool=CLIPVisionModelTaskPool/clip-vit-classification_TA8

Merge CLIP-ViT-B/32 models on eight image classification tasks using ISO-CTS, with a scaling factor of 1.5:

fusion_bench \
    fabric.loggers.name=iso_cts \
    method=isotropic_merging/iso_cts \
    method.scaling_factor=1.5 \
    modelpool=CLIPVisionModelPool/clip-vit-base-patch32_TA8 \
    taskpool=CLIPVisionModelTaskPool/clip-vit-classification_TA8

The results differ from those in the original paper due to the use of different sets of models.

[github-actions]

Images automagically compressed by Calibre's image-actions ✨

Compression reduced images by 14.9%, saving 104.11 KB.

Filename	Before	After	Improvement	Visual comparison
docs/algorithms/images/iso_merging.png	697.75 KB	593.65 KB	-14.9%	View diff

65 images did not require optimisation.

[Copilot]

Copilot reviewed 6 out of 13 changed files in this pull request and generated 3 comments.

Files not reviewed (7)

• mkdocs.yml: Evaluated as low risk
• fusion_bench/method/classification/continual_clip_finetune.py: Evaluated as low risk
• fusion_bench/dataset/fer2013.py: Evaluated as low risk
• fusion_bench/method/init.py: Evaluated as low risk
• config/method/task_singular_vector/TaskSingularVectorMerging.yaml: Evaluated as low risk
• fusion_bench/method/task_singular_vector/TSVM.py: Evaluated as low risk
• fusion_bench/method/isotropic_merging/init.py: Evaluated as low risk

Comments suppressed due to low confidence (1)

fusion_bench/method/isotropic_merging/iso_utils.py:65

• Replace print statements with proper logging.

print(f"Combining by avg {key}...")