[MM] speed up OPs using hf models (clip, ...)

[drcege]

Currently, when set np=28, clip of vit-base-p32 takes over 1h to compute similarities for 558k dataset, and tens of hours for vit-large-p14-336.

Perhaps the following can help:

1. loading on GPU (implemented)
2. using batched computing (not easy to implement, as batching is closely related to the internal logic of operators)

[github-actions]

This issue is marked as stale because there has been no activity for 21 days. Remove stale label or add new comments or this issue will be closed in 3 day.