Add managed TensorRT-LLM serve backend

[Bobcat]

Summary

• Add trtllm_serve as a managed local backend alongside llama.cpp and vLLM serve.
• Start and stop TensorRT-LLM in its own process group, wait for health, and forward chat-completion requests.
• Expose TensorRT-LLM runtime settings through model definitions and router administration.
• Add a disabled Gemma 4 26B A4B NVFP4 example with bounded compilation concurrency.
• Cover configuration parsing, lifecycle, request translation, response handling, and router dispatch.

Validation

• 197 passed, 7 skipped, 6 subtests passed
• Loaded /home/gunnar/models/nvidia/Gemma-4-26B-A4B-NVFP4 with TensorRT-LLM 1.3.0rc25.
• Completed text generation through the managed llm-pool backend.
• Verified startup health checks, graceful process-group shutdown, and the OpenAI-compatible response mapping.

Scope

The example model remains disabled by default. The repository does not install TensorRT-LLM or its CUDA/OpenMPI runtime.

MTP speculative decoding is intentionally not enabled. The current upstream shared-KV implementation depends on the trtllm-gen attention backend used on SM100/SM103, while this RTX PRO 6000 SM120 system uses FlashInfer FA2. An eager SM120 proof of concept worked, but CUDA graphs and batch sizes above one are not ready for this backend.