Refactor all of Attention CUDA kernels to have separate kernels and shared functionality

[titaiwangms]

We need a better organization and clean up on contrib op MHA and GQA, and ONNX standard Attention: attention CUDA kernels to have a better readability and maintainability.

(1) ONNX Attention needs its own kernel: QkvToContext. It is now using GQA and MHA kernel, but they have different spec, so the implementation of ONNX attention is a bit awkward.
(2) Identify duplication code and functions and then make shared functions for them (MHA/GQA/ONNX Attention) if possible.
(3) The performance prefers flash attention to memory efficient attention and then unfused attention.
(4) ONNX Attention might need 2 kernels, since nonpad_kv_seqlen makes a big difference between 23 and 24 opset version. Also, with inplace kv cache possible in TensorScatter, two modes of Attention should be considered.

The ONNX Attention:
https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/core/providers/cuda/llm/attention.cc

The ONNX Attention Spec:
https://github.com/onnx/onnx/blob/main/docs/Operators.md#Attention

The main folder of Contrib Op LLM Attention: https://github.com/microsoft/onnxruntime/tree/main/onnxruntime/contrib_ops/cuda/bert
Some important files beginning of tracing:
https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/contrib_ops/cuda/bert/group_query_attention_impl.cu
https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/contrib_ops/cuda/bert/group_query_attention.cc
https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/contrib_ops/cuda/bert/multihead_attention.cc
https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/contrib_ops/cuda/bert/attention_impl.cu
https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/contrib_ops/cuda/bert/attention_prepare_qkv.cu

Contrib Op specs: https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.MultiHeadAttention
https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.GroupQueryAttention

The core of the kernel (flash attention):
https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/contrib_ops/cuda/bert/flash_attention/flash_api.cc

[kunal-vaishnavi]

I agree with this. Note that there is a DecoderMaskedMultiHeadAttention kernel inside MHA that is not supported for GQA. That kernel is designed for beam search attention in OpenAI's Whisper.

Regarding how the attention files are organized, some are in the bert folder while others are in the llm folder. I think it would be good to have an attn folder and move all of the remaining bert files into llm. The bert name is a past relic and should be modernized.