Support I64 token mappings in safetensors model loading

[maxprogrammer007]

This PR fixes model loading for vocab-quantized models whose mapping tensor is stored as I64 in the safetensors header. The loader previously assumed all mappings were 32-bit integers and decoded the raw bytes with chunks_exact(4), which produced incorrect token ID mappings for models like potion-code-16M.

Changes

• Decode mapping based on t.dtype() instead of assuming I32
• Support both I32 and I64 mappings
• Return a clear error for unsupported mapping dtypes
• Add a regression test covering both I32 and I64 decoding

Validation

• cargo test decode_token_mapping --lib
• cargo test quantized_models_match_float32

fixes #40

[Pringled]

@maxprogrammer007 thanks for fixing this. There was a small formatting issue which caused CI to fail, I've fixed it on your branch. Will release this soon!