Add Softcapping implementation

[gkielian]

Consider adding this as a separate option inside the attention block:

https://huggingface.co/blog/gemma2#soft-capping-and-attention-implementations

From the above

Putting it all together, the logits are calculated by: logits ← soft_cap ∗ tanh(logits/soft_cap)

Gemma 2 employs soft capping for the final layer and for every attention layer. The attention logits are capped at 50.0, and the final logits at 30.0.