refactor: Migrate raw function pointers to std::function

[byronxu99]

This PR replaces several raw function pointers with std::function objects.

The biggest changes are to the learner class, which is now no longer templated. To erase the data type, learner builders will bind the data object to their input functions via a lambda capture. To erase the example type, the lambda will accept a polymorphic_ex wrapper around VW::example* and VW::multi_ex*, which implicitly converts back to the original type. The resulting lambda can be stored in a std::function object with a type which does not depend on DataT nor ExampleT.

This removes all undefined behavior around function pointer typecasting.

[rajan-chari]

Yeah there are a few ways of looking at it. You should ask a few people to see what the default intuition would be. For me it is the perspective of examples since thats what the logic of reductions is designed around

Maybe it's helpful to seperate the issues when thinking about naming things. There are 2 tensions:

1. From the stack construction perspective we move in one direction (bottom of the stack to top). When predict() is called, we move in the other direction (top of stack calls the next reduction down the stack).
2. The word 'base' has two meanings when going down the predict/learn stack. (a) It's the next down stream reduction. (b) It's the last reduction in the stack.

[rajan-chari]

Perf results seem good. It makes sense to pass shared_ptr by reference to avoid allocation, locking and copy overhead. I think most people here would have been watching out for that anyway. Looks good to me. Thanks!