fix: avoid cleanup errors for partially initialized LlamaModel

[usernames122]

This solves a bug I uncovered, that causes an AttributeError if constantly re-initializing a model in a loop and Python garbage collects it, such as testing the highest GPU layer count you can go before CUDA OOMs.
Specifically, it solves:

    if self.sampler is not None:
       ^^^^^^^^^^^^
AttributeError: 'LlamaModel' object has no attribute 'sampler'

Fixes #2104