[GPU][BUG] Incomplete conditions for sdpa prefill and head_size (#31727)

### Details:
 - *item1*
 - *...*

### Tickets:
 - *CVS-172034, CVS-169994*

---------

Co-authored-by: Chen Peter <[email protected]>

Author: luo-cheng2021

src/plugins/intel_gpu/src/graph/impls/ocl_v2/sdpa/sdpa_opt.cpp

@@ -123,7 +123,7 @@ class SDPAOptImpl : public SDPAImplBase {
         // So far this case was observed only from the non-lm models such as vision embedding model.
         // If we need to optimize unaligned head size SDPA for 2nd+ token phase of LM model,
         // we'll need to fix single_token kernel to support unaligned head size.
-        if (is_prefill || unaligned_head_size(params)) {
+        if (is_prefill || unaligned_head_size(new_params)) {
             GPU_DEBUG_TRACE_DETAIL << "execute multi_tokens for prefill with indirect = " << is_indirect << "\n";
             return execute_stage(events, instance, is_indirect ? indirect_multi_tokens : regular_multi_tokens);
         }