[onert/ggma] Support 4bit quantization for tinyllama

[glistening]

Caution

tinyllama is Maykeye/TinyLlama-v0

Currently ggma supports only float models for decode.
Let's add 4bit quantization (e.g. ggml q4).

• Write a tool to quantize f-circle to q-circle [tools/oquantize] Introduce oquantize #16291
• Add 4bit kernel for Attention Op