HF_ENDPOINT support

[jshield]

seems like the HF_ENDPOINT env var is only partially supported, It retrieves metadata via it, but models are still resolved from the huggingface endpoint; looking at download.py, the base url appears to be hardcoded.

It might help to some extent with reducing upstream burden if a local caching service like https://github.com/vtuber-plan/olah can be used as a local mirror.

#217

[FromCSUZhou]

same question