We should have a more robust check. @jcbjcbjc

we can use null lora slot?
Like:

• Reserve some slot in LoraMemPool and if lora_path is None in the loop, assign the weight_indices[i] for that request to the reserved null slot index.

And set some parameters so that applying LoRA (here rank 0) has absolutely no effect. But then backend kernel should handle this case as well?