Method comparison: Image benchmark shows sample images

[BenjaminBossan]

The Gradio space for the PEFT benchmarks already contains logic to show the image generation results once the experiments have been run and pushed to the repo. However, it doesn't contain any logic to show generated sample images. Those are quite important, as the metrics are not sufficient to estimate the quality of the fine-tuned model.

This PR adds a gallery element below the dataframe that shows sample images. By default, it just shows the training images from the dataset to give viewers an impression of what they're dealing with. If a user clicks on a row in the dataframe, e.g. the row for LoRA, the five sample images for that PEFT method are shown (if they can be found in the HF bucket). There is a toggle for the user to switch between showing generated sample images and training images.

Morever, in this PR, I made a change to use json.dumps to put the configs into the dataframe instead of simple coercing them to string. That makes it easier to parse the data.

I also noticed that in the dataframe, we would show the combined metrics between all tasks, which included metrics that are not shared. So e.g. for the image gen task, we would show test accuracy, and for MetaMath, we would show DINO similarity. The PR changes the logic to only show relevant metrics.

Moreover, the image gen metrics were on the far right of the dataframe, making them hard to view. Now the order puts them further to the left.

Finally, I added a short task description above the dataframe so that viewers can understand at a glance what they're looking at. It includes a link to the checkpoints on HF buckets.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.