Support softcap, softmax_precision in Attention(CUDA)

[titaiwangms]

Follow up #27542 and #27613

We should support

onnxruntime/onnxruntime/core/providers/cuda/llm/attention.cc

Line 1097 in dcf23fb

// TODO(titaiwang): Support softcap and softmax_precision on CUDA kernels.

Please check the previous works and the current TODO as issue description.

Attention spec also explains what we should expect to these parameters: https://github.com/onnx/onnx/blob/main/docs/Operators.md#Attention

Reference to cpu implementation: https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/core/providers/cpu/llm/attention.cc

Reference to related test files:
https://github.com/microsoft/onnxruntime/tree/main/onnxruntime/test/python/transformers/test_onnx_attention
https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/test/providers/cpu/llm/attention_op_test.cc
https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/test/testdata/onnx_backend_test_series_filters.jsonc

[titaiwangms]

qk_matmul_output support is deferred because it's more complicated and more of a debugging helper only.

onnxruntime/onnxruntime/core/providers/cuda/llm/attention.cc

Line 1175 in dcf23fb

// TODO(titaiwang): Support additional output_qk modes beyond kNone and kQK.