AMDGPU EP: register umbrella plugin EP and forward profile options

[zz002]

Summary

Makes the AMDGPU umbrella EP work end-to-end through OGA's own model_benchmark with profile=hip. Three OGA-side fixes in src/amdgpu; no umbrella EP change required.

Why

Testing with OGA's bundled C model_benchmark (which has no --ep_library/-p/--profile flag; profile is passed via genai_config.json provider options) surfaced three gaps:

1. Plugin EP not registered. The AMDGPU umbrella is a plugin EP, so its OrtEpDevice is only discoverable after RegisterExecutionProviderLibrary. Legacy in-proc EPs skip this, but the umbrella does not, so AppendExecutionProvider_V2 finds no device and allocation fails ("no requested allocator available"). The RyzenAI interface already self-registers its DLL; the AMDGPU path did not.
2. Profile not reaching the umbrella. The umbrella reads its provider options (notably profile) from session-config entries prefixed ep.amdgpuexecutionprovider., not the AppendExecutionProvider_V2 ep_options channel, so profile=hip never selected the hipgpu backend.
3. Trivial init session had no profile. EnsureDeviceOrtInit builds a trivial device-init session to set up the device allocator. After the Route init-session provider-option shaping through DeviceInterface microsoft/onnxruntime-genai#2232 refactor, its provider options come solely from the per-EP ShapeInitSessionProviderOptions hook (default no-op), so the umbrella had no backend to create for that session.

What

1. EnsureUmbrellaEpRegistered() in src/amdgpu/session_options.cpp — self-registers amdgpu-ep.dll on the OrtEnv, mirroring the RyzenAI pattern (resolve the DLL next to the genai/ort module or the executable; re-registration is benign).
2. ForwardUmbrellaProviderOptions() in src/amdgpu/session_options.cpp — bridges the genai_config provider options into ep.amdgpuexecutionprovider.* config entries so the umbrella selects the requested backend.
3. ShapeInitSessionProviderOptions() override in src/amdgpu/interface.cpp — forwards the user's provider options to the trivial init session so it carries the profile too.

Test plan

• Built model_benchmark.exe + onnxruntime-genai.dll locally with the umbrella amdgpu-ep.dll + hip-backend.dll.
• profile=hip runs end-to-end on Llama-3.1-8B (int4-awq) via model_benchmark.exe on gfx1151 (StxHalo); prefill + decode complete, output correct.
• Perf within ~3% of the 2194 baseline (TTFT/TPS); no regression from the added registration/forwarding.

Notes for reviewers

• Umbrella EP allocator memory-info name is irrelevant to allocator lookup (matched by device attributes: GPU + AMD vendor id + device index + memory type), so no umbrella-side change is needed here.
• StxHalo default-heuristic enablement in the umbrella EP is a separate follow-up.