Extra lora fix

[daanelson]

We have a bug right now in our hotswap loras where if a users passes a given extra_lora for a prediction, and then removes that extra_lora but otherwise doesn't change lora_scale or replicate_weights, the extra_lora will persist.

Any change to lora_scale, replicate_weights, or adding a new extra_lora should reset the weights (assuming that request maps to the instance that has the extra_lora that's sticking around) - but still, this is the very definition of a weird, annoying bug.

The fix below explicitly monitors for changes in extra_lora and handles it the same way that changes in lora_weights are handled. It also handles the case where extra_lora_weights are passed in and lora_weights aren't; this is an edge case, but good to handle.