feat(client): backport TensorRT-LLM support to release/0.5.1

[zhengluo-nv]

Summary

Backport the following merged changes to release/0.5.1:

• 03923943aa779b9b4bfee5290c6b819a1ce1e444 — native TensorRT-LLM P2P example (feat: add native TensorRT-LLM P2P example #540)
• 3542dda4c5cb96c533b299eb4b034dce93755782 — TensorRT-LLM engine adapter (feat(client): add TensorRT-LLM engine adapter #565)
• 6175ec3ca768a1b0bab84ae1efd4480af8b62168 — protobuf 6 runtime support (fix(client): support protobuf 6 runtimes #564)

The commits are cherry-picked in dependency order with -x provenance. The adapter conflict resolution preserves the release branch's newer source-peer cleanup behavior while adding exact tensor-catalog matching and the dynamic NIXL import path. Its test was adapted to the release branch's already-removed InstantTensor strategy.

Impact

This brings the native TensorRT-LLM integration and example to the 0.5.1 release line and allows the Python client to run with protobuf 5 or 6.

Validation

• 64 passed — focused TensorRT-LLM, NIXL, transfer-safety, RDMA peer-release, and protobuf tests with protobuf 5.29.6
• 32 passed — core TensorRT-LLM/NIXL/protobuf tests with protobuf 6.33.6
• git diff --check
• All three backport commits have valid GPG signatures and matching DCO sign-offs

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!