Faster LoRA Loading

[daanelson]

This enables faster loading of LoRAs for fine-tuned models. Changes:

• Pass tuned lora weights in at predict time
• Cacheing weights locally for improved inference w/the same model.