Enable zero-copy for QNN GPU

[qti-mattsinc]

• Use the new GPU shared memory allocator in QNN EP to
  allocate the KV cache on CPU-accessible GPU memory.
  This provides a large speedup by eliminating unnecessary
  copy overhead.
• Refactor QNN-specific checks out of the common
  EnsureDeviceOrtInit by adding GetMemoryInfo and
  GetProviderOptionsForAllocatorSession to DeviceInterface.

[johnpaultaken]

This change does not look ideal to me, we need to discuss why it deviates from the norm.
Ideally what I like to see is no change to genai at all.
When OrtMemoryInfo is of type QnnShared can we just return the QnnGpuAllocator based on the device selection made by the user ie GPU ?
Everything else should work transparently, just like how it works for the other EPs cuda, openvino etc.
I don't see a need for enable_dx12_shared_memory_allocator option, why not enable it always in the EP ?
Also we cannot have variable names like use_dx12_shared_memory etc which then turns out to be only a Qnn specific option. Ideally, I would want to avoid any code in genai that are EP specific.

[Copilot]

Pull request overview

Enables zero-copy KV-cache placement for the QNN GPU backend by using the shared-memory allocator exposed by recent QNN EP packages, with a runtime capability probe and fallback. Also refactors device-specific allocator/session-option logic out of the common EnsureDeviceOrtInit path by adding two new virtual methods (GetMemoryInfo, GetProviderOptionsForAllocatorSession) on DeviceInterface, with QNN and WebGPU providing overrides.

Changes:

• Add DeviceInterface::GetMemoryInfo / GetProviderOptionsForAllocatorSession plus a DeviceTypeToString helper, and replace ad-hoc per-device branches in EnsureDeviceOrtInit with calls to these virtuals.
• Add IsQNNGPUBackend and a IsQNNGPUSharedAllocatorAvailable runtime probe (embedded trivial ONNX model) to opt the GPU backend into shared-memory allocation when supported, with a warning-and-fallback otherwise.
• Move WebGPU's "WebGPU_Buf" vs "WebGPU_Buffer" name fallback into the WebGPU GetMemoryInfo override.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/smartptrs.h	Adds DeviceTypeToString and two new virtual methods on DeviceInterface with default implementations driven by per-device name tables.
src/models/model.cpp	Simplifies EnsureDeviceOrtInit to use the new virtuals; drops inline QNN/WebGPU special cases and uses DeviceTypeToString in error messages.
src/qnn/interface.h	Forward declarations plus IsQNNGPUBackend declaration.
src/qnn/interface.cpp	QNN overrides of the new virtuals (with GPU vs NPU device filtering) and implementation of IsQNNGPUBackend.
src/qnn/session_options.cpp	Adds shared-memory allocator capability probe via a hardcoded trivial ONNX model; selects QNN device only when GPU shared allocator is available or NPU shared allocator is explicitly enabled.
src/webgpu/interface.cpp	Adds WebGPU GetMemoryInfo with attempted fallback between "WebGPU_Buf" and "WebGPU_Buffer".

[Copilot]

Pull request overview

Copilot reviewed 7 out of 7 changed files in this pull request and generated 7 comments.

[johnpaultaken]

Looks cleaner than existing code.
Qnn specific checks are moved to qnn specifie interface now.
And so is web gpu specific code.