feat(isolation): DynamicVRAM compatibility for process isolation

DynamicVRAM's on-demand model loading/offloading conflicted with  process isolation in three ways: RPC tensor transport stalls from mid-call GPU offload, race conditions between model lifecycle and active RPC operations, and false positive memory leak detection from changed finalizer patterns.

- Marshal CUDA tensors to CPU before RPC transport for dynamic models
- Add operation state tracking + quiescence waits at workflow boundaries
- Distinguish proxy reference release from actual leaks in cleanup_models_gc
- Fix init order: DynamicVRAM must initialize before isolation proxies
- Add RPC timeouts to prevent indefinite hangs on model unavailability
- Prevent proxy-of-proxy chains from DynamicVRAM model reload cycles
- Add torch.device/torch.dtype serializers for new DynamicVRAM RPC paths
- Guard isolation overhead so non-isolated workflows are unaffected
- Migrate env var to PYISOLATE_CHILD

comfy/k_diffusion/sampling.py

@@ -192,7 +192,7 @@ def sample_euler(model, x, sigmas, extra_args=None, callback=None, disable=None,
     """Implements Algorithm 2 (Euler steps) from Karras et al. (2022)."""
     extra_args = {} if extra_args is None else extra_args
     s_in = x.new_ones([x.shape[0]])
-    isolation_active = args.use_process_isolation or os.environ.get("PYISOLATE_ISOLATION_ACTIVE") == "1"
+    isolation_active = args.use_process_isolation or os.environ.get("PYISOLATE_CHILD") == "1"
     if isolation_active:
         target_device = sigmas.device
         if x.device != target_device: