fix(search): guard CrossEncoderReranker.ensure_initialized with asyncio.Lock

[Srujan-rai]

Summary

• Two concurrent requests that both observed _initialized=False would both
  call initialize(), double-loading the local model and causing a memory spike
• Added a lazily-created asyncio.Lock with double-checked locking pattern so
  only the first caller runs initialization; subsequent callers wait and return
  immediately once _initialized=True

Tests

Added tests/test_reranker_init_lock.py:

• test_concurrent_ensure_initialized_calls_initialize_once — fires 3 concurrent
  ensure_initialized() calls via asyncio.gather, asserts initialize called once
• test_second_call_after_init_skips_initialize sequential second call is a no-op

Test plan

• Concurrent recall requests no longer double-initialize the reranker model
• Memory usage stable under concurrent load with local cross-encoder