[GPU] Improve sdpa_opt kernel performance with flashattn2 softmax tricks. #40987

Job	Run time
Smart_CI	22s
Docker	47s
Build / Build	5m 59s
Debian Packages	0s
OpenVINO JS API	0s
ONNX Runtime Integration	0s
Samples	0s
CPU functional tests	0s
PyTorch Models tests	0s
OpenVINO tokenizers extension	0s
C++ unit tests / C++ unit tests	40s
Python unit tests / Python unit tests	2m 18s
Pytorch Layer Tests / PyTorch Layer Tests	2m 34s
JAX Layer Tests / JAX Layer Tests	1m 20s
Python API tests / Python API tests	0s
TensorFlow Models tests	0s
TensorFlow Layer Tests	0s
ci/gha_overall_status_linux_arm64	1s
	14m 1s

Provide feedback