[GPU] Fixed wrong idx for applying attn_sink in pa_sdpa_opt (#32731)

### Details:
 - Sink to be appiled to each query idx instead of head_idx 
- Currently there is no practical situation this code is activated.
However potentially there will be, e.g., igpu + sink + pa

### Tickets:
 - *ticket-id*

Author: yeonbok

src/plugins/intel_gpu/src/graph/impls/ocl_v2/paged_attention_opt.cl

@@ -399,7 +399,7 @@ KERNEL(pa_sdpa_opt)(
             GET_VECTOR_ELEMENT(exp_sum, q_idx) = sub_group_reduce_add(GET_VECTOR_ELEMENT(exp_sum, q_idx));
             #ifdef HAS_SINK_INPUT
             const uint head_idx = get_global_id(1);
-            GET_VECTOR_ELEMENT(exp_sum, head_idx) += (native_exp(TO_SOFTMAX_ACCUMULATOR_TYPE(sink_ptr[head_idx] - GET_VECTOR_ELEMENT(qk_max, q_idx))));
+            GET_VECTOR_ELEMENT(exp_sum, q_idx) += (native_exp(TO_SOFTMAX_ACCUMULATOR_TYPE(sink_ptr[head_idx] - GET_VECTOR_ELEMENT(qk_max, q_idx))));
             #endif
         }