GLM-5.1: drop --detokenizer-worker-num 4 (multi-detokenizer broken in dev-cu12)

The dev-cu12 sglang build crashes immediately on the first chat request
when --detokenizer-worker-num >1 is set:

  File 'multi_tokenizer_mixin.py', line 494, in event_loop
    ipcs is not None
  AssertionError: Batch req recv_obj.rids=['...'] has invalid http_worker_ipcs
  SIGQUIT received.

Reverts that one flag to default (1). Keeps --chunked-prefill-size 8192
and --watchdog-timeout 600 since those are plain config tweaks and the
watchdog increase still helps the detokenizer-falling-behind pattern.

Author: Evrard-Nil

GLM-5.1.yaml

@@ -126,7 +126,6 @@ services:
       --model-loader-extra-config '{"enable_multithread_load": "true", "num_threads": 64}'
       --enable-mixed-chunk
       --chunked-prefill-size 8192
-      --detokenizer-worker-num 4
       --watchdog-timeout 600
       --port 8000
       --host 0.0.0.0