feat(openengine): add sibling server for TRT-LLM

[connorcarpenter15]

Summary

• add an optional OpenEngine gRPC sibling server to trtllm-serve while preserving the existing HTTP and native gRPC interfaces
• share one LLM instance across protocols with protocol-neutral admission, cancellation, draining, request tracking, and exactly-once shutdown
• implement generation, discovery, health, load, abort, drain, LoRA lifecycle, KV connector discovery, and KV event/statistics streaming
• support aggregate and context-first prefill/decode roles, ordered image/video/audio media, per-request media options, guided decoding, logprobs, priority, cache controls, and lazy LoRA selection
• fan out TRT-LLM's single-consumer KV and statistics queues so HTTP and OpenEngine do not compete

Why

This provides an engine-neutral process boundary for Dynamo without replacing TRT-LLM's current in-process, HTTP, or native gRPC integrations. The same TRT-LLM engine can serve HTTP and an OpenEngine sidecar endpoint in one process.

Protocol dependency

• feat(protocol): add revision 2 multimodal contract and packages ai-dynamo/openengine#20
• pinned OpenEngine schema commit: cea19cb06acf03c911b84d5c147e519b60dd92a6

Validation

• TensorRT-LLM wheel built successfully in the TRT-LLM development container
• OpenEngine unit and fake-servicer suite: 64 passed
• targeted context-first PEFT/LoRA prefill-decode test passed
• native gRPC compatibility, conversion, cancellation, drain, fanout, LoRA, and shutdown coverage passed
• computelab frontend smokes passed for Qwen3 aggregate text/LoRA, Qwen3 1P1D text/LoRA, Qwen3-VL image/video aggregate and 1P1D, and Phi-4 aggregate audio

Compatibility

OpenEngine is disabled unless --openengine-port is supplied. Existing HTTP behavior remains enabled, and the dedicated native --grpc mode is rejected only when combined with the new sibling server.