Improve generate embeddings function by using dataset tools

[michel-aractingi]

What this does

Refactors generate_embeddings.py to use dataset tools. Specifically, modify_features to add the image and language embeddings and remove videos in one call. Dataset tools also writes parquet files in a more optimized way.

How it was tested

Tested by running generate embeddings script on this dataset aractingi/utokyo_embeddings

I ran

python src/lerobot/datasets/generating_embeddings/generate_embeddings.py \
        --repo-id lerobot/utokyo_xarm_bimanual \
        --output-repo-id aractingi/utokyo_embeddings \
        --image-encoder dinov2_vitb14 \
        --language-encoder minilm-l12 \
        --remove-videos \
        --push-to-hub

Then :

python src/lerobot/datasets/generating_embeddings/validate_embeddings.py \
        --original-repo-id lerobot/utokyo_xarm_bimanual \
        --embeddings-repo-id aractingi/utokyo_xarm_bimanual_embeddings \
        --image-encoder dinov2_vitb14 \
        --language-encoder minilm-l12 \
        --num-samples 10

Both passed