Use lazily-evaluated coroutines for expression templates

[pfeatherstone]

Have you guys thought about using coroutines for the implementation details of an expression template?
You can lazily co_await tensor elements from previous operations and co_yield tensor elements for the current implementation.
The compiler could then hopefully see the entire "state machine" of the combined operation and lower that to something optimal.
Maybe that would work?

More crucially, for users needing to define custom operations in terms of an expression template, rather them customize some complicated class, they just define the operation in terms of a coroutine using loops.