Static attention: support local-global attention (pytorch#13043)

Summary:

Runtime: support different cache lengths for different layer.
Python: add sliding window cache update which was already in the runtime.

Reviewed By: billmguo

Differential Revision: D79267644

Author: sxu