[CPU] PagedAttention supports dynamic-split fuse

[luo-cheng2021]

Details:

• Merge first token and second token inference into one parallel loop
• ~~Additional optimization: pre-transpose k-cache, pre-pack v-cache if needed~~
• Additional optimization for first token: save q * k' upper triangle matrix computation and (q * k') * v lower triangle matrix computation
• C++ pipeline can enable it: https://github.com/ilya-lavrenov/openvino.genai/pull/4
• TODO(in another PR):
  • alibi support
  • performance tuning
  • testcase

Tickets:

• 138673