Add Gemma 3 examples (4B, 12B, 27B)

[EwoutH]

Hi,

Cerebras fast inference is amazing, however the model selection is a bit limited. I would suggest adding some variants of Gemma 3. On many tasks they are right on the cost-performance pareto frontier:

The Gemma 3 models are allowed to be commercially offered via API services. I think having fast interference of the 4B, 12B, 27B would be extremely valuable.

[EwoutH]

Notably, the Gemma 3 QAT models perform almost identically with a third to a quarter of the memory usage: