[Fix] Bind ring drain workers to their CUDA device

[XbzOnGit]

Closes #105.

Summary

Each native ring and D2H stream is created on the model worker's current CUDA device, but the background drain thread previously started with logical device 0 current. Under tight GPU-memory pressure, this caused the entire rank-1 capture to fail bit-for-bit validation—all 35,840 rank-1 rows failed—while model inference itself still completed.

This change records the owning device before creating the drain stream and selects that device at worker entry before any CUDA operation. A startup handshake makes cudaSetDevice() failure visible synchronously to the caller.

It also promotes failures from D2H enqueue, drain-stream synchronization, and the worker's final device synchronization from debug-only output to release-build diagnostics. This is diagnostic hardening only; transactional D2H failure handling is outside this PR.

Regression coverage

The new two-GPU native regression blocks work on device 0, constructs the ring and drain on device 1, and verifies that stopping the drain does not wait for device 0. This deterministically checks the worker's device binding without depending on GPU-memory pressure.

Validation

• Native ring suite: 52 passed, 0 failed.
• Qwen3-0.6B, BF16 eager, TP=2, vllm-full, vLLM GPU-memory utilization 0.96, 512 MiB payload and pinned-staging rings:
  • rank 0: 87,296 / 87,296 exact rows passed;
  • rank 1: 35,840 / 35,840 exact rows passed;
  • total: 123,136 / 123,136 exact rows passed.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

Suppressed comments (1)

tests/native/ring/test_ring_engine.cu:285

• The wait loop for callback_entered also busy-spins for up to 5 seconds. Adding a small sleep reduces CPU load in CI while still polling frequently enough for this test.

    while (!callback.entered.load(std::memory_order_acquire) &&
           std::chrono::steady_clock::now() < callback_deadline) {
        std::this_thread::yield();
    }

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

[zaoxing]

LGTM