Add use_exact_model_name option to prevent automatic model name modification

[niryuu]

When loading quantized models, Unsloth automatically modifies the model name to load optimized versions. While this is helpful in most cases, it can lead to duplicate model caching when users specifically want to load both original and quantized ways.

This PR adds a new use_exact_model_name parameter that allows users to bypass this automatic modification and load the exact model specified.

Example:

model, tokenizer = FastLanguageModel.from_pretrained(
    "google/gemma-2-9b",
    load_in_4bit=True,
    use_exact_model_name=True  # Will load exactly this model without modification
)

[danielhanchen]

So the goal is instead of loading google/gemma-2-9b-bnb-4bit, it should load directly from the cache of google/gemma-2-9b?

[niryuu]

Yes. Unsloth has a mechanism for rewriting the given model name according to unsloth/models/mapper.py for efficiency. However, there are cases where we want to use the exact model name, such as cache control. The purpose of this PR is to provide an option for this case.