Add WebGPU INT2 and mixed-width QMoE support (#32878)
## Description
Adds WebGPU execution support for INT2 and mixed-width integer `QMoE`,
following the operator contract introduced for CUDA in #32761.
- Accepts 2-bit expert weights on WebGPU.
- Passes FC1, FC2, and FC3-specific bit widths to the existing
`MatMulNBits` implementation.
- Replaces the mixed-width rejection test with a nonzero INT2 FC1 / INT4
FC2 numerical test for the optimized single-token path.
The underlying WebGPU `MatMulNBits` shaders already implement 2-bit
unpacking, so no new shader path is required.
## Validation
- [x] `clang-format --dry-run --Werror` on changed C++ files
- [x] `git diff --check`
- [x] VS Code diagnostics clean
- [x] Release `onnxruntime_provider_test` build with WebGPU/D3D12
- [x] `QMoETest_WebGPU_MixedWidth_SingleToken` in hardware-backed WebGPU
CI
- [x] All WebGPU build/test CI jobs passed
Local virtual D3D12 execution is unstable with
`DXGI_ERROR_DEVICE_REMOVED`; hardware-backed CI is the numerical
validation source.
## Review follow-up (October 5)
Added H200/Vulkan coverage for homogeneous inherited INT2, k=2
mixed-width indirect selection of two distinguishable experts, two-token
mixed-width routing to different experts, and separate non-fused SwiGLU
FC3 with an independent INT2 width. FC1, FC2, and optional FC3 now
reject packed rows/blocks that cannot be indexed as u32 words; tests
cover the 12-element INT2 boundary and FC3. All 16 focused WebGPU QMoE
tests pass locally; all six review threads have replies and are
resolved. CI is rerunning on the new head. The previous Android minimal
binary-size check exceeded its 1,597,440-byte threshold by 509 bytes;
this is not claimed fixed by the WebGPU-only changes and must be
reassessed on the rerun.
---------
Co-authored-by: Sayan Shaw <sayanshaw@microsoft.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
