Fix Engine atexit cleanup lifetime

[galitheia]

Motivation

Composer currently registers the same static _close callback for every
Engine and passes each Engine's state and logger into atexit. Manually
closing an Engine does not unregister that entry, retaining training objects
and allowing cleanup to run again during interpreter shutdown.

Additionally, Engine.__del__() unregisters the shared callback globally,
which can remove handlers belonging to other live Engines. This prevents
bounded repeated in-process Trainer lifetimes.

This change gives every Engine a unique weak atexit callback, makes close()
idempotent, and unregisters only the callback owned by that Engine.

Summary

• Register a unique weak atexit callback for each Engine.
• Make Engine.close() idempotent and unregister only that Engine's callback.
• Preserve the interpreter-shutdown sentinel and cleanup ordering.
• Delegate del cleanup through close().
• Add coverage for independent Engines, idempotent close, and exactly-once
  interpreter-exit cleanup.

Testing

• Direct Engine lifecycle and interpreter-exit probes passed.
• Ruff lint passed.
• YAPF formatting check passed.
• git diff --check passed.