Make MultiHeadAttention to work with all attention operators

[WenjieDu]

1. Feature description

Since #333, we have made self-attention operator replaceable in Transformers. Although there're other attention operators in pypots, e.g. ProbAttention from Informer, but not all of them can work with MultiHeadAttention. So we need a refactor.

2. Motivation

This can increase code reusability and make things in pypots more like Lego。

3. Your contribution

Will make a PR to achieve this.