[codex] Bump vLLM requirement to 0.21

[Fr0do]

Summary

• Bump the plugin vLLM dependency requirement from >=0.19.0 to >=0.21.0.
• Update the README requirement line to match.

Validation

• Ran the reproducible SR004 setup script: /workspace-SR004.nfs2/kurkin/scripts/setup_mera_t5gemma2_vllm018_env.sh.
• Ran the MERA/MMRED one-sample T5Gemma2 smoke script successfully: /workspace-SR004.nfs2/kurkin/scripts/run_mera_t5gemma2_smoke.sh.
• MLflow run: /workspace-SR004.nfs2/kurkin/mlruns/581881116757115999/0547b279c7b345d284ff37f386a3c515.

Caveat

• The SR004 node currently has NVIDIA driver 560.35.03; the separate vLLM 0.21.0 CUDA probe refuses to install on this driver to avoid breaking the working environment. The successful smoke validated the plugin path using the existing SR004 vllm 0.18.0 stack, not a full vLLM 0.21.0 runtime.