[codex] Remove vLLM layerwise reload alias patch

[samsja]

Summary

• Remove the prime-rl monkey patch for vLLM layerwise reload alias buffers.
• Stop overriding vLLM's _copy_and_restore_kernel_tensors during the general plugin setup.

Details

vLLM 0.22.0 already registers MambaMixer2.conv_weights as a non-persistent buffer and its layerwise reload metadata skips non-persistent buffers that alias parameter storage. Keeping the prime-rl patch conflicts with that behavior because the original kernel tensor list can still contain conv_weights while the temporary restored layer intentionally does not.

This returns the checkpoint-format /update_weights path to vLLM's own reload implementation.

Validation

• UV_NO_SYNC=1 uv run ruff check src/prime_rl/inference/patches.py
• Local vLLM reload probe after transformers_v5_compat():
  • verified layerwise._copy_and_restore_kernel_tensors remains the original vLLM function
  • verified MambaMixer2.__init__ registers conv_weights with persistent=False
  • verified vLLM's default reload metadata omits a non-persistent child-parameter alias buffer and restores the layer without AttributeError or parameter corruption
• Runtime validation on local stack with [codex] Fix external-LB inference config sizing #2705 + this PR, Slurm job 23624:
  • all 4 API servers reached Application startup complete
  • orchestrator initialized NCCL with 4 servers, inference_world_size=16, gpus_per_server=4
  • all 16 inference workers logged Reloading checkpoint-format weights with vLLM layerwise processing
  • all 4 /update_weights admin requests returned 200 OK
  • no Enabled vLLM layerwise reload alias-buffer patch, conv_weights, AttributeError, Internal Server Error, or Traceback appeared in the runtime logs
  • W&B run: https://wandb.ai/primeintellect/nemotron_sami_debug/runs/e2f73ef0996346cf9c4fe91732a4796f