[DO NOT MERGE]: swap vLLM wheel to upstream v0.21.0 for LoRA + EP

[JohannesHa]

The PrimeIntellect custom wheel (0.21.0+cu129.r42434.pr39568.a106aa6) was branched off vLLM main before PR #40867 ("Initial EP support for LoRA", merged 2026-05-09) landed, so it still asserts against base_layer.use_ep in FusedMoEWithLoRA and rejects MoE LoRA + EP. Upstream v0.21.0 (cut 2026-05-15) carries the merge and serves EP + LoRA cleanly. Verified locally on 2 GPUs with Qwen3-30B-A3B + jeeejeee/qwen3-moe-text2sql-spider at tp=2, EP=true.

Side effect: the routed-experts / NIXL P/D path that the fork wheel carried is no longer available, so prime-rl's
enable_return_routed_experts reads are guarded with getattr to keep the patches/serving-tokens imports working on the upstream wheel.

---

Note

Medium Risk
Swapping the x86_64 vllm wheel from a custom fork to upstream may change inference/runtime behavior and performance, especially around P/D (NIXL) and LoRA/EP paths. The code changes are small but sit in inference patching/serving logic and could affect routed-experts response shaping.

Overview
Moves x86_64 installs off the PrimeIntellect custom vllm wheel to the upstream v0.21.0 CUDA 12.9 wheel (with corresponding uv.lock updates), primarily to pick up upstream LoRA + expert-parallel support.

Updates prime-rl’s routed-experts/NIXL integration to be compatible with upstream vLLM by guarding enable_return_routed_experts reads with getattr(..., False) in both the vLLM config __post_init__ patch and the /inference/v1/generate non-streaming response post-processing path.

^{Reviewed by Cursor Bugbot for commit 80137f3. Bugbot is set up for automated code reviews on this repo. Configure here.}

[JannikSt]

great job