GH-62052: add defer_close logic to FileIO

[nascheme]

If the FileIO object is wrapped by a write buffer, set the defer_close property to true. That ensures the underlying FileIO object is not closed until the buffer is closed (and can flush the unwritten data). This avoids data loss, as demoed by buffer_not_flushed.py. E.g. if both FileIO and the BufferedWriter are both part of the reference cycle, the finalize triggered by the GC can close the FileIO object before BufferedWriter has a chance to write the data out.

This is a proof-of-concept at this point, not ready to merge.

• Issue: Built-in module _io can lose data from buffered files in reference cycles #62052

[nascheme]

Some further thoughts about this:

• defer_close should be a counter, rather than a bool. We would need to keep track of the number of write buffers using the FileIO object. Only close it when the (unflushed?) buffer count reaches zero.
• checking buffer count in the internal_close function is probably not the right thing. If someone calls FileIO.close() then we could just close and possibly lose data. If count > 0 can be checked on GC finalize, I think that would be sufficient to fix the Built-in module _io can lose data from buffered files in reference cycles #62052 bug.
• perhaps if count > 0 and close is called, the FileIO should remember that. Then when count reaches zero the file is automatically closed at that time. Essentially, remember that close was called but don't do it until buffers have a chance to flush.
• the buffer count is a bit like a special kind of reference count, keeping track of # of write buffers using the FileIO object