Fixed Quantization bug in TransformerLens 3.0

[jlarson4]

Description

While reviewing #684, I noticed that for BnB-4bit Mistral: q_proj.weight, k_proj.weight, v_proj.weight, o_proj.weight are all Params4bit with dtype torch.uint8 (packed). Bridge picks the first param's dtype as the "compute dtype" and casts hidden_states from fp32 to uint8, which clamped all values to [0, 255] and truncated decimals. HF's MistralAttention.forward then runs on this destroyed input and produced gibberish.

By ensuring this type cast on hidden_states is not done with integer dtypes, we get bit-precise results when compared against HuggingFace. Also added regression tests that will fail loudly if this cast is ever unintentionally restored for Quantized models.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility