Close instruments when the daemon is done with them, not in __del__

[sequoiap]

Fixes #61. This is independent of #98: no shared files, so they can merge in either order.

Problem

Instrument.__del__ called close(), which caused three problems:

• The base class's close() raises NotImplementedError, and many drivers' close() fails on an instrument that never connected (self.device doesn't exist).
• Errors raised in __del__ go to stderr, and a daemon runs in a windowless child process, so nobody sees them.
• An instrument in a reference cycle (threads, callbacks) is only freed when Python's garbage collector next runs, which may be never before the process exits.

Change

• When the daemon stops: DaemonRunner.run closes every single-mode instance it created, through Daemon.release_instruments(). This happens on the error path too.
• When a client disconnects: Daemon.clientDisconnect closes that client's session-mode instances. Pyro discards them as soon as the connection closes anyway. LockableDaemon.clientDisconnect now chains to it after releasing locks.
• Instrument._release() handles every call:
  • it calls close() at most once, so shutdown followed by garbage collection doesn't close an instrument twice;
  • it skips classes that don't override close();
  • it logs errors as warnings with the traceback, instead of raising them.
• __del__ remains a last resort through the same path, logging at debug level and never raising.
• Docs: the base close() docstring and CLAUDE.md now say when close() is called, and that it must cope with an instrument that never connected.

Not covered: a session-mode instance whose client is still connected when the daemon stops. That still depends on __del__, which is now harmless. Covering it would mean going through Pyro's internal connection list for each server type.

Things to know

• Drivers whose close() fails on an instrument that never connected (ppcl55x, the RTO, KDC101 and others) will log a warning at shutdown if their instrument was never connected. Before, that error went to an invisible stderr. Each of those close() methods can be made tolerant in a later PR.
• Instruments now really are closed when a daemon stops. Before, that only happened if garbage collection got to them before the process exited.

Testing

• test_service.py:
  • _release closes an instrument once, leaves a class without close() alone, and logs errors;
  • garbage collection never raises. This test fails on master: the base class's __del__ reports an unraisable exception.
• test_server.py, end to end over real sockets:
  • disconnecting closes that client's session instruments only, with both Daemon and LockableDaemon;
  • single-mode instruments survive disconnects and are closed by release_instruments().
• test_registration.py: a real DaemonRunner with an instrument in a reference cycle, and the garbage collector paused. Stopping the runner closes the instrument. On master it is never closed.
• Full suite: 442 passed, with pydantic's deprecation warnings turned into errors.

🤖 Generated with Claude Code