Revert NVFP4 over-quantization of the MTP routed experts. Restores the source-format MTP shard (model-00046-of-00046.safetensors: I8 weight + F8_E8M0 scale), removes the 768 mtp.* entries from hf_quant_config.json quantized_layers and adds mtp.* to exclude_modules, and corrects the stale total_size in model.safetensors.index.json. Only these 3 files change; all other shards are untouched.
