[GPU] Optimize Eltwise

[byungilm]

• Fusing Eltwise to Broadcast
• Implemented fusing operation to broadcast_ref kernel

Description of the issue

• From analysis of Qwen-Reranker (CVS-173218), Eltwise occupied 6% execution time with ref kernel.
  it can be optimized out.

The code and line that caused this issue

• Modified broadcast ref kernel : kernel_selector/cl_kernels/broadcast_gpu_ref.cl
• Added condition for fusing Eltwise to Broadcast : graph_optimizer/prepare_primitive_fusing.cpp
• Added logic to fused post-ops : broadcast/broadcast_kernel_base.cpp

Reproduction step and snapshot

• target model QWen3-Reranker-0.6B is in openvino_notebook (notebooks/Qwen3-embedding)
• Reproduced by benchmark_app
  ./benchmark_app -m openvino_notebooks/notebooks/qwen3-embedding/Qwen3-Reranker-0.6B/FP16/openvino_model.xml -shape [64,256] -d GPU -hint latency -api sync -nireq 1 -niter 1

Checklist

• Is it a proper fix?
• Did you include test case for this fix, if necessary?
• Did you review existing test that can be extended to cover this scenario?

Tickets:

• CVS-173226