Enable HIP backend routing for Strix Halo LLM workloads

[urpetkov-amd]

Enable HIP backend routing for Strix Halo LLM workloads

Summary

This change enables the AMDGPU umbrella/plugin Execution Provider to select the
HIP backend for LLM workloads on Strix Halo hardware.

HIP is selected only when all of the following conditions are true:

1. The project is built with USE_HIP=ON.
2. The detected GPU architecture starts with gfx1151.
3. The AMDGPU provider option model_arch is llm.
4. The selected profile is Auto.

All other hardware, model architectures, profiles, and builds without HIP
support retain their existing routing behavior.

Behavior matrix

Build	Hardware	model_arch	Auto result
USE_HIP=ON	gfx1151	llm or case variant	HIP
USE_HIP=ON	gfx1151	missing or another value	MIGraphX
USE_HIP=ON	another gfx11* device	llm	MIGraphX
USE_HIP=OFF	gfx1151	llm	MIGraphX
Any	Any	Any, with model_fw=webnn	DirectML
Any	Any	Any, with an explicit profile	Explicit profile

Runtime limitation

End-to-end HIP inference was not completed because hipgpu.dll was not
available in the local environment.

hip-backend.dll is a thin loader built by this repository. The actual HIP EP
implementation is provided by the external hipgpu.dll:

amdgpu-ep.dll -> hip-backend.dll -> hipgpu.dll -> amdhip64_7.dll

The external DLL must be supplied by the HIP EP artifact/package and placed in
the runtime DLL search path, normally beside hip-backend.dll. This is
separate from the ROCm HIP runtime DLL amdhip64_7.dll.