feat(kv-offload): use external PegaFlow server

[xiaguan]

Closes #724

Depends on novitalabs/pegaflow#407.

Summary

• replace the embedded PegaFlow engine with an external CUDA IPC/RPC client
• keep host, SSD, RDMA, and routing configuration in the PegaFlow server; OpenInfer exposes only the endpoint and content namespace
• support Qwen3 and GLM5.2 arena registration with opt-in exportable CUDA allocations
• carry block stride inside each CUDA IPC tensor and fail closed on mixed registration-wire revisions
• use terminal native Load completion and flush submitted saves before unregistering the client instance

Validation on final PR heads

• prek run
• cargo test --release -p openinfer-kv-offload --lib (9 passed)
• PegaFlow server tests (23 passed) and core tests (124 passed, 1 ignored)
• PegaFlow server all-targets and Python client CUDA 13 checks

Cross-process gate status

The Qwen3 byte-exact restore and GLM5.2 EP8 host-restore gates passed before the PegaFlow PR reached its final scope. Those results remain design evidence, but the gates must be replayed on the final heads before deployment.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4d4f98b04e

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Check release RPC status before dropping leases

When a lease release reaches the server but is rejected at the application layer, this call currently drops the RPC response and returns Ok(()); unlike health/save/load/flush, it never validates the returned status. In the reservation-pressure and abandoned-prefetch paths this makes failed releases look successful, so host blocks can remain pinned until the server TTL instead of being surfaced as a release failure.

Useful? React with 👍 / 👎.