llama.cpp Adds MiMo V2.5 Support

PR #22493 in ggml-org/llama.cpp adds text-to-text inference support for MiMo V2.5(+ Pro). The audio and vision configurations for the non-Pro model are excluded from this scope.

• The author has not yet re-verified the Pro model, but expects conversion and quantization to work correctly.
• convert_hf_to_gguf.py was fixed to correctly dequantize FP8 safetensors.
• MiMo's TP-aware sharding and fused attention_qkv were split apart again to align with the existing MiMo V2 Flash path.
• attention_value_scale from config.json was also wired into the GGUF conversion path to reduce discrepancies with the transformers implementation.

In Q8_0 verification, Mean PPL(Q) 5.135221 was nearly identical to the base 5.128919, and Mean KLD 0.012455 showed small quantization error.

In the comments, a conversion error occurred due to a missing GGUFWriter.add_attn_value_scale, but the author added a fix commit to address it.