feat: AMD GPU umbrella execution provider integration

[Wang, Zhenze (zz002)]

Summary

Integrates the AMD GPU umbrella execution provider (amdgpu-ep, a plugin EP shipped by the ROCm hip-ep project) into GenAI on top of rel-0.15.1, adding DeviceType::AMDGPU and a device interface that registers and selects the EP through the plugin EP (V2) API.

What

src/amdgpu/ holds the new device interface. It resolves the EP library from AMDGPU_EP_PATH, then next to onnxruntime-genai.dll, onnxruntime.dll and the executable, before falling back to the working directory, and registers it on the interface's OrtEnv (tolerating an already-registered library from an earlier init cycle). SetupProvider picks the EP devices reported as GPUs and appends the EP with the model's provider options; the same options are mirrored to ep.amdgpu.* session config entries, which is how the umbrella EP selects its backend. ShapeInitSessionProviderOptions forwards those options to the device-init session so the allocator can be created there, and GetMemoryInfo() builds the memory info with CreateMemoryInfo_V2 using the AMD vendor id and GPU device type the EP factory registers with.

The umbrella EP's allocator is host- and device-accessible on the AMD APU iGPU, so model inputs stay in device memory and the device is on the continuous-decoding list.

Status

This is a draft opened as a patch source for AMD's downstream CI, which currently builds GenAI from rel-0.15.1 plus this branch. It overlaps in intent with #2165 (AMD GPU EP support); once that lands and covers the umbrella EP's provider-option plumbing, this PR will be closed and the downstream CI switched over. It supersedes #2194, which targeted v0.14.0.

Test plan

• Windows build against ORT 1.27 (--no_telemetry, Ninja, Release)
• model_benchmark on Llama-3.1-8B (AWQ int4) with "AMDGPU": {"profile": "hip"} and session.disable_cpu_ep_fallback=1: matches the v0.14.0 + Add AMDGPU execution provider support #2194 baseline on time-to-first-token and token generation throughput

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.