Fix capture asymmetry in suspend_global_capture(in_=True)

[FazeelUsmani]

Fixes #13322.

If a plugin hook suspends out/err and then calls suspend_global_capture(in_=True) (e.g. to read input), the matching resume_global_capture() was incorrectly re-enabling out/err even though the caller had them suspended. Fix is to remember what was started before the nested suspend and only resume those.

Includes a pytester regression that runs the flow from a pytest_terminal_summary hook and checks that the print after the resume reaches stdout before the 1 passed line. Before the fix it only shows up at teardown, after the summary.

[RonnyPfannschmidt]

at first glacne this looks good, but also makes the age and the state-managment issues of the apis apparent

i'll need to do a deeper review

[RonnyPfannschmidt]

we should use a literal or pivot to a enum

[FazeelUsmani]

Went with a Literal, matching _CaptureMethod right above it. _assert_state takes it too, so a typo'd state now fails mypy.

[FazeelUsmani]

@RonnyPfannschmidt Thanks for taking a look. Agreed, the fix is local. The underlying issue is that suspend/resume infer state from whichever streams happen to be active at call time rather than tracking it explicitly. No rush on the deeper review. If a broader rework feels like the right path, happy to scope that into a separate issue.