Fix SuppressStdoutStderr permanently silencing the process when contexts overlap - #3290

SuppressStdoutStderr used to duplicate fd 1 and 2 in __init__ and restore them in __exit__, per instance, with no synchronization. If a second context was created while a first one was active (two threads fitting Prophet models side by side, or just overlapping with blocks), the second one saved the devnull descriptors as if they were the real ones, and whoever exited last pointed fd 1 and 2 back at devnull for the rest of the process lifetime — the scenario from #3276, where after two overlapping fits everything printed afterwards was gone. Creating the context in interpreters without usable standard streams (pythonw on Windows, GUI-embedded hosts) also raised OSError: Bad file descriptor straight from __init__, leaking the descriptors opened just before the failure.

The redirection is now shared state: a module-level lock plus a depth counter, where the outermost context captures and restores the real descriptors, contexts entered while a redirection is already active just join it, and all descriptor work happens in __enter__ with a graceful no-op fallback when the standard streams cannot be saved. I left Prophet's suppression path as-is — it becomes safe through the context manager itself, and switching it over to cmdstanpy logger levels felt like a separate design call.

How I checked it: on master, the new threaded test fails exactly like the issue describes (assert b'' == b'survivor' — the write after both contexts lands nowhere) and the fallback test dies with OSError: [Errno 9] Bad file descriptor out of __init__. With the fix both pass, along with the round-trip and nested tests and the rest of test_logging.py (8 passed). The overlap choreography is event-driven rather than sleep-based, so the interleaving is deterministic. ruff check and ruff format (0.7.2, same as pre-commit) are clean.

Closes #3276.