onnxruntime

79d8fdad - [WebGPU] Handle absent optional bias in convolution and Attention (#33719)

[WebGPU] Handle absent optional bias in convolution and Attention (#33719) ### Description Conv and ConvTranspose can receive an explicit empty optional-bias slot. WebGPU treated that slot as a tensor because it contributed to the input count, then dereferenced null. Attention similarly forwarded an omitted projection bias as a null third MatMul input. Determine bias presence from the actual tensor in convolution paths, and append Attention's MatMul bias only when supplied. Keep the positive runtime regressions WebGPU-specific: omitted, explicit-empty and nonzero supplied biases, fed/initializer weights and biases, 2D/3D convolution, FP32/FP16 spatial cases, and Attention. Explicitly select WebGPU with Full validation and disable CPU fallback. Retain the common Conv/ConvTranspose test rejecting missing required weights with an explicit bias during graph validation. Other EP fixes are deferred from this PR. ### Validation - Repository lintrunner and whole-PR whitespace checks pass. - MSVC syntax checks pass for both changed test files, including the WebGPU fixture with contrib operators disabled. - The current native rebuild stopped while writing a precompiled header because E: ran out of disk space. Fresh runtime validation of this test relocation is pending CI. - Historical Windows x64 Release / NVIDIA D3D12 validation at `413060e4ae`: all six focused regressions and 65 existing convolution tests passed with shader-cache validation. Earlier fail-before checks reproduced the WebGPU access violation without the production fixes. These results precede the latest merge and test relocation. Local native tests use a focused provider-test target with bundled Dawn. Intel-specific im2col execution, shared-plugin runtime and the full provider suite were not run locally.