Raise a clear error for QLoRA + vllm_mode="server" instead of a cryptic vLLM crash

[DaoyuanLi2816]

What does this PR do?

Rejects QLoRA with vllm_mode="server" before weight synchronization. Merging a 4-bit PEFT adapter dequantizes its weights, which cannot be loaded by a server expecting packed 4-bit tensors.

The same guard is applied to VLLMGeneration and the self-contained OnlineDPOTrainer. The error directs users to an unquantized server or colocate mode. This is a fail-fast fix for #4973, not full quantized-server support.

Tests

• server-mode guard tests for both implementations (3 passed, 1 skipped without local 4-bit support)
• Ruff check and format check on the changed files

AI writing disclosure

• AI-assisted: some parts were suggested or improved by AI, but the PR was written and reviewed by a human.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit e597239. Configure here.}