Optimizing BART: encoder/decoder attention

Nobody has claimed this yet.

**Is your feature request related to a problem? Please describe.**
I'd like to optimize BART with ONNX Runtime, but it looks like the only Attention operator currently supported is [self-attention](https://github.com/microsoft/onnxruntime/blob/master/docs/ContribOperators.md#com.microsoft.Attention), and BART requires encoder/decoder cross-attention.

**System information**
- ONNX Runtime version (you are using): 1.4.0

**Describe the solution you'd like**
A fused operator implementing encoder/decoder cross-attention

**Describe alternatives you've considered**
- In the meantime I'm planning to check out APEX's [Fast Multihead Attention](https://github.com/NVIDIA/apex/tree/master/apex/contrib/multihead_attn).

Start by reading ONNX Runtime's existing com.microsoft.Attention operator documentation and the BART encoder/decoder cross-attention requirement described here. Confirm whether the requested fused operator is still needed, then identify the relevant operator implementation and tests before defining completion criteria for cross-attention support.

Written by the indexing model from the issue text.