Fix drafter initialization and preserve draft weights

[vx120]

Fix: Ensure drafter initialization before first serving request

Summary

Fixes the issue where online drafter weights would be lost when the serving worker experiences a level-2 sleep/wake-up cycle. Implements proper state tracking and initialization flow for speculative decoding drafters.

Changes

1. Initial Drafter Initialization (_speco_ensure_initial_draft_weights)

   • Load base drafter from checkpoint before admitting first request
   • Prevent weight loss during the cold-start phase using async lock

2. Weight Source Tracking (_speco_draft_weight_source)

   • Track whether current drafter weights come from "checkpoint" or "online" publication
   • Preserve weight source during level-2 snapshot/restore cycles
   • Validate state transitions with runtime checks

3. Wake-up Recovery

   • Add fallback when online drafter snapshot is unavailable
   • Restore checkpoint-based weights and validate restoration
   • Handle uninitialized drafter states gracefully

Affected Files

• verl_speco/integration/vllm_runtime.py: Main implementation

[tpx818]

1.测试一下rollout开多DP的情况；
2.部分会触发vllm sleep level=1的情况下，online train时，权重恢复是否有问题