refactor: eliminate circular import by separating model management functions

- Create model_management_mgpu.py for centralized model lifecycle tracking
- Move memory management functions from device_utils.py to new module:
  * multigpu_memory_log, track_modelpatcher, trigger_executor_cache_reset
  * check_cpu_memory_threshold, prune_distorch_stores, try_malloc_trim
  * force_full_system_cleanup
- Update imports across codebase (distorch_2.py, distorch.py, __init__.py,
  nodes.py, checkpoint_multigpu.py)
- Resolves device_utils.py ↔ distorch_2.py circular dependency
- Follows established clean coding patterns with fail-fast error handling

Addresses critical CPU memory leak investigation infrastructure by ensuring
proper module separation for comprehensive memory management utilities.

__init__.py

@@ -12,6 +12,8 @@ from .device_utils import (
     get_device_list,
     is_accelerator_available,
     soft_empty_cache_multigpu,
+)
+from .model_management_mgpu import (
     trigger_executor_cache_reset,
     check_cpu_memory_threshold,
     multigpu_memory_log,