fix(core): instantiate each sub-client exactly once in _spawn_completion_context

Previously the first sub-client was constructed twice — once for
`other_backend_client` and again inside the registration loop. With
backends that hold expensive underlying state (HTTP connection pools,
async clients, OS sockets), the duplicate carried real cost; with
multi-backend Anthropic on macOS this showed up as a ~3-min hang at
iter 2 of every multi-backend trajectory. Usage tracking also merged
via dict.update, so the duplicate overwrote one of the two instances'
usage summaries.

Reuse `other_backend_client` for the first registration and only
construct the remaining sub-clients in the loop.

Author: alexzhang13

rlm/core/rlm.py

@@ -206,9 +206,16 @@ def _spawn_completion_context(self, prompt: str | dict[str, Any]):
 
         lm_handler = LMHandler(client, other_backend_client=other_backend_client)
 
-        # Register other clients to be available as sub-call options (by model name)
-        if self.other_backends and self.other_backend_kwargs:
-            for backend, kwargs in zip(self.other_backends, self.other_backend_kwargs, strict=True):
+        # Register other clients to be available as sub-call options (by model name).
+        # Reuse other_backend_client for the first entry so each (backend, kwargs)
+        # pair is instantiated exactly once.
+        if other_backend_client is not None:
+            lm_handler.register_client(other_backend_client.model_name, other_backend_client)
+            for backend, kwargs in zip(
+                self.other_backends[1:],
+                self.other_backend_kwargs[1:],
+                strict=True,
+            ):
                 other_client: BaseLM = get_client(backend, kwargs)
                 lm_handler.register_client(other_client.model_name, other_client)