refactor visit_Output

refactor visit_Output

* `finalize` is generated once and cached for all nodes.
* Extract common behavior for native env.

Removed the compiler behavior where groups of nodes would generate a
format string. Instead, individual nodes are always yielded. This made
rendering 30% faster in the examples, and simplifies the code. It also
removes the issue where Python would report either the first or last
line of the multi-line format expression, messing up the traceback line
number mapping.