Retire the GPU workers when a run that held a card finishes

A worker that has run a jax node keeps holding the GPU after the run: XLA
preallocates most of the VRAM at import and never releases it, so the next
process OOMs on preallocation while a warm idle worker sits on the card.
Pools are kept warm on purpose — a library reads its environment at import,
so a warm worker cannot be re-told — but the end of a run is a point where
the memory should go back, and the environments carrying a GPU assignment
are exactly the pools that ran on one. Idle ones go now, busy ones when
they return.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019Hra4ndWMCLU5F3KjUuVAc

backend/fluksio/flow/workers.py

@@ -321,6 +321,24 @@ class PythonWorkerPool:
         for child in children:
             child.respawn_all()
 
+    def retire_gpu_children(self) -> None:
+        """Retire the pools holding a card, so the VRAM goes back.
+
+        A library like JAX takes most of the device when it imports and never
+        releases it, so a warm worker that has run one such node is a held
+        card — and warm is the point of a pool, so nothing retires it. At the
+        end of a run there is something to key on: the environments carrying a
+        GPU assignment are exactly the pools that ran on one.
+        """
+        with self._lock:
+            children = [
+                child
+                for key, child in self._children.items()
+                if any(name == "CUDA_VISIBLE_DEVICES" for name, _ in key)
+            ]
+        for child in children:
+            child.respawn_all()
+
     def _drain(self) -> list[_Worker | None]:
         slots = []
         while True: