feat: enhance parameter annotation resolution in task signatures

[solvemproblr]

Problem

@env.template breaks hud deploy / sync introspection when env files use from __future__ import annotations and template params use rich types (Literal, aliases, Pydantic models, unions).
Failure surfaces as TaskArgs is not fully defined.

Cause

_TaskFactory stored a raw inspect.signature(func).
With postponed annotations, that signature keeps annotation strings, and those unresolved strings were consumed by both manifest schema generation and start-time arg coercion.

Solution

Use an evaluated signature at construction time:

• In _TaskFactory.__init__, changed to inspect.signature(func, eval_str=True).
• Existing callers keep using task.sig, but now they get real runtime types.
• This fixes both manifest args schema generation and runtime coercion without adding new signature APIs.

---

Note

Low Risk
Single-line signature evaluation change with targeted tests; improves typing/introspection without altering wire protocol or auth paths.

Overview
Fixes hud deploy / sync failures (TaskArgs is not fully defined) when env modules use from __future__ import annotations and @env.template parameters use rich types (Literal, aliases, Pydantic models, unions).

_TaskFactory now stores inspect.signature(func, eval_str=True) instead of a raw signature, so task.sig exposes resolved runtime types for both tasks.list args JSON Schema (_args_json_schema) and wire tasks.start arg coercion (_coerce_args in TaskRunner).

Adds tests that postponed annotations produce correct manifest enums/$ref/defaults and that a full Run coerces string wire args into typed template parameters.

^{Reviewed by Cursor Bugbot for commit a6ac044. Bugbot is set up for automated code reviews on this repo. Configure here.}

[jdchawla29]

lgtm