gh-116738: Make zlib module thread-safe

[yoney]

Makes the zlib module thread-safe free-threading build. Even though operations are protected by locks, attributes exposed via PyMemberDef (eof, needs_input, unused_data, unconsumed_tail) should still be stored atomically within locked sections, since they can be read without acquiring the lock.

• Replaced PyThread_type_lock with PyMutex (similar to gh-116738: Use PyMutex for bz2 module #140555 and gh-116738: Use PyMutex for lzma module #140711)
• Removed ENTER_ZLIB/LEAVE_ZLIB macros in favor of direct PyMutex_Lock/Unlock (similar to gh-116738: Use PyMutex for bz2 module #140555 and gh-116738: Use PyMutex for lzma module #140711)
• Used FT_ATOMIC_STORE_CHAR_RELAXED for eof and needs_input flag updates inside locked sections (attributes exposed via PyMemberDef)
• Used _PyObject_XSetRefDelayed for unused_data/unconsumed_tail inside locked sections

cc: @mpage @colesbury

• Issue: Audit all built-in modules for thread safety #116738

[colesbury]

I think we should avoid using _PyObject_XSetRefDelayed for these attributes and use a lock instead (maybe self->mutex if possible).

_PyObject_XSetRefDelayed is useful when data is read-mostly, but I don't think that's the case for unused_data or unconsumed_tail. I'm also concerned that _PyObject_XSetRefDelayed() will keep potentially delay freeing large amounts of memory.

[yoney]

we should avoid using _PyObject_XSetRefDelayed

I think we have a situation similar to the one below, and if we want to avoid using _PyObject_XSetRefDelayed() in Thread A, I might move unused_data and unconsumed_tail from PyMemberDef to PyGetSetDef and lock/unlock self->mutex in Thread B. Is there an easier or better way to do this?

Thread A                                         Thread B (read via PyMemberDef)
═══════════════════════════════════════════════|═══════════════════════════════════════════
[LOCK self->mutex]                             |            
...                                            |
Py_SETREF(self->unused_data, new_unused_data); | FT_ATOMIC_LOAD_RELAXED(self->unused_data)
...                                            |             
[UNLOCK self->mutex]                           | 

[emmatyping]

unused_data is almost entirely read-only. It's only read (usually once) after decompression has completed, and written once after flush() is called.

unconsumed_tail on the other hand is written to by the decompressor pretty much every decompress() call.

[colesbury]

I might move unused_data and unconsumed_tail from PyMemberDef to PyGetSetDef and lock/unlock self->mutex in Thread B. Is there an easier or better way to do this?

This makes sense to me.