A node whose FUNCTION points at a method that does not exist (e.g. a typo in
a custom node), or a V3 node missing its execute override, was only detected
once that node ran -- after every upstream node had already executed. In a
multi-node workflow the user waited for the whole graph to run up to the
broken node before seeing the error.
validate_prompt already walks every node before execution; add an
executability check there so the error is reported up front and attributed
to the offending node (returned in node_errors), and nothing runs.
The check resolves the V1 FUNCTION method on the class (the common case) and
falls back to an instance, since the runtime invokes it on an instance and a
node may define FUNCTION or its method in __init__. V3 nodes are checked via
their existing VALIDATE_CLASS.
Add tests for V1 typo, V3 typo, good nodes, and a node whose method is
defined in __init__ (must not be falsely rejected).
