fix vram estimation when loading a 16bit model with load_in_4bit=True

[yizhe-skyfall]

No description provided.

[mmathew23]

Hi @yizhe-skyfall thank you for the contribution. I'm not able to replicate the issue on my side. For example if I put a debug statement in approximate_vllm_memory_usage in the current version of unsloth_zoo, I see that it print out true when loading in 4bit and fast_inference = True. Perhaps the issue was addressed since you first raised the PR. Could you verify if all is working as expected for you?