[webgpu] Fix multihead-attention for ort-web-tests (#24485)

Compute 'total_sequence_length' the same way as JSEP.

Author: jchen10

onnxruntime/contrib_ops/webgpu/bert/attention.cc

@@ -446,7 +446,8 @@ Status ApplyAttention(const Tensor* Q, const Tensor* K, const Tensor* V, const T
                       WebgpuAttentionParameters& parameters, onnxruntime::webgpu::ComputeContext& context, const Tensor* seqlen_k) {
   const int output_count = std::min({context.OutputCount(), 1 + (past_key != nullptr ? 1 : 0) + (past_value != nullptr ? 1 : 0)});
   const int past_sequence_length = output_count > 1 ? parameters.past_sequence_length_ : 0;
-  const int total_sequence_length = parameters.total_sequence_length_;
+  const int total_sequence_length =
+      parameters.is_gqa_ ? parameters.total_sequence_length_ : past_sequence_length + parameters.kv_sequence_length_;
 
   const TensorShapeVector probs_dims({parameters.batch_size_, parameters.num_heads_,
                                       parameters.sequence_length_, total_sequence_length});