transformers

8817bc56 - Add the block-scaled quantizer base and the fine-grained modules

Add the block-scaled quantizer base and the fine-grained modules BlockScaledQuantizer holds what the block-scaled formats share: the fine-grained linear, grouped linear and experts modules, their load and save ops, the scale layouts, the device requirements and the dequantize fallback. Each format quantizer subclasses it. FineGrainedHfQuantizer can host them. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>