Fix async task review regressions

[D3f0]

Problem

Post-merge review of PR #121 reproduced three correctness regressions in async task support:

1. hide=True leaks stdout/stderr from the async runner.
2. Synchronous tasks in an invocation containing an async task receive a coroutine from ctx.run() and silently skip their command.
3. Grouped pre/post task entries can crash async-task detection before execution.

The review also found an under-annotated public ToolkitContext.run surface and documented async-path limitations.

Acceptance criteria

• Async hide=True suppresses both terminal streams while retaining captured Result output.
• Only coroutine task bodies execute inside async_task_context; synchronous graph members retain blocking ctx.run() semantics.
• Async-task detection recursively handles grouped pre/post entries; execution does not treat grouping containers as tasks.
• ToolkitContext.run exposes the synchronous-or-awaitable Result return contract to type checkers.
• Async documentation removes resolved limitations and retains the current buffered-output, non-interactive/watchers, sudo, and sync-helper caveats.
• Focused async/grouped-task tests cover each regression.

Reference: docs/async-review-findings.md.