fix(core): drain GPU workers before unregister

[xiaguan]

Summary

• reject new GPU transfers and drain both worker streams before removing an instance
• synchronize CUDA streams at the drain barrier, including transfer error paths
• make server cleanup wait for the engine drain before releasing registered CUDA tensors
• serialize cleanup against registration and session creation, and keep cleanup running after caller cancellation
• preserve the synchronous unregister APIs while adding drained variants for allocation owners

Testing

• prek run
• core unit tests: 127 passed, 1 ignored
• server unit tests: 23 passed
• HTTP cleanup integration test: 1 passed

[xiaguan]

Folded into #407 so native CUDA IPC registration and the required drain-before-unmap lifecycle are reviewed and merged together.