[WebGPU] Correct MatMul bias input indexing in MatMul::ComputeInternal

[Copilot]

Description

WebGPU MatMul was wiring bias into the wrong slot in the inputs vector when bias is present. The vector is pre-sized to 3 and downstream code reads bias from inputs[2], but the previous logic appended bias, placing it at index 3.

• Root cause
  • inputs was initialized with size 3 for bias cases, then bias was added via push_back, creating a 4th element.
• Change
  • Replaced append logic with direct assignment to the expected index.
• Effect
  • Bias is now consistently passed at inputs[2], matching ComputeMatMul() and Intel-path expectations.

std::vector<const Tensor*> inputs(has_bias ? 3 : 2);
inputs[0] = a;
inputs[1] = b;
if (has_bias) {
  const auto* bias = context.Input(2);
  inputs[2] = bias;  // was: inputs.push_back(bias);
}

Motivation and Context

This addresses incorrect bias propagation in WebGPU MatMul (onnxruntime/core/providers/webgpu/math/matmul.cc) caused by index mismatch between producer and consumer logic. The fix is intentionally surgical: one behavioral line change in the affected code path.