Add logit softcapping to GQA

[kunal-vaishnavi]

Description

This PR adds the softcap attribute to the GroupQueryAttention op.

Motivation and Context

This PR helps resolve the NaN output issue with Gemma-2 raised in this issue.

[fxmarty-amd]

ORT 1.19.2 does not support softcap. I later had:

onnxruntime_genai.onnxruntime_genai.OrtException: Load model from /home/fxmarty/repos/onnxruntime-genai/src/python/py/models/llama_oga/model.onnx failed:This is an invalid model. In Node, ("/model/layers.0/attn/GroupQueryAttention", GroupQueryAttention, "com.microsoft", -1) : ("/model/layers.0/attn/qkv_proj/MatMulNBits/output_0": tensor(float),"","","past_key_values.0.key": tensor(float),"past_key_values.0.value": tensor(float),"/model/attn_mask_reformat/attn_mask_subgraph/Sub/Cast/output_0": tensor(int32),"/model/attn_mask_reformat/attn_mask_subgraph/Gather/Cast/output_0": tensor(int32),"cos_cache": tensor(float),"sin_cache": tensor(float),) -> ("/model/layers.0/attn/GroupQueryAttention/output_0": tensor(float),"present.0.key": tensor(float),"present.0.value": tensor(float),) , Error Unrecognized attribute: softcap for operator GroupQueryAttention

[kunal-vaishnavi]

The softcap attribute was added here. This PR has been merged to prepare for the stable releases of ONNX Runtime v1.20 and ONNX Runtime GenAI v0.5.0, which will be happening soon. You can build ONNX Runtime GenAI from source with a nightly ONNX Runtime version by copying over the necessary files to ort for now until the releases occur.

[fxmarty-amd]

@kunal-vaishnavi Thank you, makes sense