[js/webgpu] support FlashAttention-2 for attention operator

[xhcao]

Description

Motivation and Context

[xhcao]

The FlashAttention-2 algorithm is based on the paper https://tridao.me/publications/flash2/flash2.pdf.
Memory increase quadratically in the sequence length in original algorithm, and need a lot of global memory read and write accesses. If sequence is large, for example running stable diffusion 2.1 with attention nodes, original algorithm will lead chrome crash for out of memory and inefficient.
Please take a look, @axinging @hujiajie @jiechen0826

[guschmue]

/azp run ONNX Runtime Web CI Pipeline,Windows GPU CI Pipeline,Linux Android Emulator QNN CI Pipeline

[guschmue]

/azp run Linux CPU CI Pipeline,Linux CPU Minimal Build E2E CI Pipeline,Linux GPU CI Pipeline,Linux GPU TensorRT CI Pipeline,Linux OpenVINO CI Pipeline,Linux QNN CI Pipeline,MacOS CI Pipeline,Windows ARM64 QNN CI Pipeline,Windows CPU CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[guschmue]

/azp run Windows GPU TensorRT CI Pipeline,onnxruntime-binary-size-checks-ci-pipeline,orttraining-linux-ci-pipeline,orttraining-linux-gpu-ci-pipeline,orttraining-ortmodule-distributed,Windows x64 QNN CI Pipeline,Big Models

[azure-pipelines]

Azure Pipelines could not run because the pipeline triggers exclude this branch/path.

[guschmue]

/azp run Windows GPU CUDA CI Pipeline,Windows GPU DML CI Pipeline,Windows GPU Doc Gen CI Pipeline

[azure-pipelines]

Azure Pipelines could not run because the pipeline triggers exclude this branch/path.

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[xhcao]

Sorry for the error, I wanted to keep the names as the paper, but the variables‘ names mismatched rules. I had already modified the names. Thanks.