fix: evaluate task annotations consistently across Python versions

[tobiasraabe]

Annotation evaluation used different namespaces before Python 3.14, and ordinary signature inspection could evaluate lazy annotations before their saved locals were available. A task whose annotation references a deleted definition-local name could collect successfully and then fail during execution.

Use one namespace-building path across Python versions and read signatures without evaluating annotations throughout collection, parametrization, and execution. Default arguments supply missing names without overriding module globals or builtins. Regression tests cover deleted local bindings, default recovery, global/default collisions, explicit namespaces, and current module globals.

Validation:

• 114 annotation, collection, generator, and decorator tests passed on Python 3.13 and 3.14.
• just lint and just typing passed.

This is the first PR in a stack. Follow-up #1026 enables definition-local snapshots for future annotations and releases generated-task snapshots after collection hooks finish. Merge this PR first, then retarget #1026 to main.