transformers

f7378ff5 - Simplify GGUF renaming with regex alternations; minimise diff vs main

Simplify GGUF renaming with regex alternations; minimise diff vs main * Consolidate per-arch renames using (q|k|v|o) / (weight|bias) / (gate|up|down) alternations — T5 drops from 29 to 19 rules, etc. * Revert all unrelated edits in `core_model_loading.py`, `base.py`, and `quantizer_finegrained_fp8.py` so the diff vs main is GGUF-only. * GGUF state-dict entries are now :class:`GGUFQuantizedTensor` (a torch.Tensor subclass carrying quant_type), so the standard loader flows them through .to(device) untouched and the :class:`GGUFDequantize` op — injected at the head of every WeightConverter chain by `GGUFQuantizer.update_weight_conversions` — dequantises on-device. Same pattern as Fp8Dequantize + Fp8Quantizer.