Consider Cythonizing the Function evaluation code

[brandonwillard]

We might be able to speed up Aesara's compiled function evaluations by Cythonizing classes like Function—or at least the methods within it and/or the functions it calls.

Every time an aesara.function-compiled graph is evaluated Function.__call__ is called, so, if we can speed that up, we can reduce the inherent overhead incurred when repeatedly calling compiled graph functions. Note that Type methods like Type.filter are called repeatedly within this process, as well as methods within aesara.link.basic.Container, so this effort may reach out into other core classes as well, since we will likely want to Cythonize those in order to reduce the in-and-out-of-Python overhead.

Regardless, it's worth considering how much Cythonization we could get away with through the entire codebase, especially since we can (theoretically) maintain pure Python code that's Cythonizable by way of annotations and .pxd files (i.e. pure Python mode).