fix(actor): harden lifecycle, supervision, Redis mailbox, and add comprehensive tests

- Fix spawn() zombie cell: clean up registry on start() failure
- Fix shutdown(): cancel + await tasks that exceed graceful timeout
- Fix _shutdown(): await mailbox.close() to release backend resources
- Fix escalate directive: stop failing child before propagating to grandparent
- Fix RedisMailbox.put(): wrap Redis errors in try/except, return False on failure
- Fix retry.py: replace assert with proper raise for last_exc
- Add put_batch() to Mailbox abstraction for single-roundtrip bulk enqueue
- Add RedisMailbox.put_batch() with atomic Lua script for bounded queues
- Add MailboxFullError exception type for semantic backpressure handling
- Add redis>=7.4.0 dependency with public PyPI sources in uv.lock

Tests added (31 total, up from 27):
- test_middleware_on_restart_hook: verifies middleware.on_restart() on supervision restart
- test_ask_propagates_actor_exception: ask() re-raises original exception type
- test_ask_propagates_exception_while_supervised: exception propagates; root actor survives
- test_ask_timeout_late_reply_no_exception: late reply after timeout is silent no-op
- test_actor_backpressure.py: MailboxFullError + dead letter on full mailbox
- test_actor_retry.py: ask_with_retry with exponential backoff
- test_mailbox_redis.py: RedisMailbox put/get/batch/close
- bench_actor_redis.py: RedisMailbox throughput benchmarks

backend/tests/bench_actor.py

@@ -2,7 +2,6 @@
 
 import asyncio
 import time
-import statistics
 
 from deerflow.actor import Actor, ActorSystem, Middleware
 
@@ -17,7 +16,11 @@ class CounterActor(Actor):
         self.count = 0
 
     async def on_receive(self, message):
-        self.count += 1
+        if message == "inc":
+            self.count += 1
+            return self.count
+        if message == "get":
+            return self.count
         return self.count
 
 
@@ -69,6 +72,8 @@ async def bench_tell_throughput(n=100_000):
         await ref.tell("inc")
     # Wait for all messages to be processed
     count = await ref.ask("get", timeout=30.0)
+    if count != n:
+        print(f"  warning: expected {n} processed, got {count}")
     elapsed = time.perf_counter() - start
 
     await system.shutdown()