fix: prevent orphan VIT model processes on crash or interrupt

[sufubao]

Summary

• Add start_parent_check_thread() to VIT model worker processes so they self-terminate if the visual server dies, matching the pattern used by router, audio, and other server components
• Return process handles from start_model_process() and store them in VisualManager and VisualOnlyManager
• Implement clean_up() to kill tracked VIT processes on shutdown, with race-condition-safe exception handling

Test plan

• Deploy visual server with VIT model processes and verify they start correctly
• Kill the visual server process and confirm VIT workers self-terminate within ~10 seconds
• Trigger an initialization error and verify clean_up() terminates all spawned VIT processes
• Verify no orphan VIT processes remain after repeated crash/restart cycles

[gemini-code-assist]

Code Review

This pull request introduces process tracking and cleanup for VIT model processes within the visual server. It updates the model startup sequence to return process handles, adds a parent process check thread, and implements a cleanup method. The review feedback suggests optimizing the cleanup logic in both VisualManager and VisualOnlyManager by signaling all processes to terminate before joining them, which prevents cumulative delays during shutdown.

[gemini-code-assist]

The current cleanup logic kills and joins processes sequentially. If multiple processes are unresponsive, the cumulative timeout (5 seconds per process) could significantly delay the shutdown of the visual server. It is more efficient to signal all processes to terminate first, and then join them.

Suggested change

	for proc in getattr(self, "model_procs", []):
	try:
	if proc.is_alive():
	logger.info(f"Killing VIT model process {proc.pid}")
	proc.kill()
	proc.join(timeout=5)
	except (ProcessLookupError, OSError):
	pass
	procs = getattr(self, "model_procs", [])
	for proc in procs:
	try:
	if proc.is_alive():
	logger.info(f"Killing VIT model process {proc.pid}")
	proc.kill()
	except (ProcessLookupError, OSError):
	pass
	for proc in procs:
	try:
	proc.join(timeout=5)
	except (ProcessLookupError, OSError):
	pass

[gemini-code-assist]

Similar to the VisualManager, the cleanup logic here is sequential. Signaling all processes to terminate before joining them would improve shutdown efficiency, especially when dealing with multiple model processes.

        procs = getattr(self, "model_procs", [])
        for proc in procs:
            try:
                if proc.is_alive():
                    logger.info(f"Killing VIT model process {proc.pid}")
                    proc.kill()
            except (ProcessLookupError, OSError):
                pass
        for proc in procs:
            try:
                proc.join(timeout=5)
            except (ProcessLookupError, OSError):
                pass