feat: Add simpler HuggingFace repository support for lora_url

[zsxkib]

This update lets users easily use LoRA models from HuggingFace by just providing the repository name (like 'owner/repo') in the lora_url parameter.

New features:

• Support for simple HuggingFace format ('owner/repo')
• Smart prioritization - tries Replicate first, then HuggingFace for owner/model format
• Clear logging when multiple model files are found
• Better model type detection from LoRA filenames
• Keeps working with all existing URL formats:
  • Direct URLs
  • Replicate URLs
  • Replicate delivery URLs

Key implementation details:

• We carefully preserve backward compatibility by always checking for Replicate models FIRST
• The 'owner/repo' format is only treated as HuggingFace after Replicate lookup fails
• Helpful error messages show when a format is tried and falls back to another
• Model type detection (1.3b vs 14b) checks filenames to pick the right base model
• Zero changes to existing direct URL handling - they continue to work exactly as before

Technical additions:

• Added HuggingFace Hub API for repository file listing
• Strong error handling and fallbacks
• Better documentation
• No changes to replicate_weights.py
• Minimal changes overall - only what's needed for the new feature

Tested with test suite:

sudo cog predict \
  -i 'model="14b"' \
  -i 'frames=17' \
  -i 'prompt="MY_SUBCONSCIOUS monster is swimming ..."' \
  -i 'lora_url="fofr/wan-14b-my-subconscious"' # replicate model

sudo cog predict \
  -i 'model="14b"' \
  -i 'frames=17' \
  -i 'prompt="MY_SUBCONSCIOUS monster is swimming ..."' \
  -i 'lora_url="https://replicate.com/fofr/wan-14b-my-subconscious"' # replicate url

sudo cog predict \
  -i 'model="14b"' \
  -i 'frames=17' \
  -i 'prompt="MY_SUBCONSCIOUS monster is swimming ..."' \
  -i 'lora_url="https://replicate.delivery/xezq/qocyf3gptqQwJqBXaeMVS8kESLBmeIjvnqrFboh1CNigawvoA/trained_model.tar"' # replicate delivery url

sudo cog predict \
  -i 'model="14b"' \
  -i 'frames=17' \
  -i 'prompt="steamboat willie style, golden era animation ..."' \
  -i 'lora_url="https://huggingface.co/benjamin-paine/wan-lora/resolve/main/steamboat-willie-14b.bf16.safetensors"' # hf url

sudo cog predict \
  -i 'model="14b"' \
  -i 'frames=17' \
  -i 'prompt="steamboat willie style, golden era animation ..."' \
  -i 'lora_url="benjamin-paine/wan-lora"' # hf repo