Use q3f16 models

[bil-ash]

@CharlieFRuan From the docs, it can be seen that mlc supports q3f16 quantization . However, all the models used in web-llm are either q4 or q0. Is q3f16 quantization not supported in web-llm(only supported in mlc-llm)? I an asking this because if q3 is indeed supported in web-llm it should be used because generally mobile browsers are memory constrained.