Mang/eagle3 similarity

[mangguo321]

Details:

• Root cause: EAGLE3 needs to reorder and compact the KV cache after it selects the valid tokens from a speculative token tree.
  Previously, PaKVReorder changed the Key and Value cache tensors in place, but exposed only a dummy output tensor. In other words, the graph did not explicitly say that the updated K/V caches were results of the operation.
  Because of this, the runtime could execute the reorder using temporary internal buffers. The reorder operation modified those buffers, but the updated data was not guaranteed to be written back to the external KV cache used by the next decoding step.
  As a result, EAGLE3 could select the correct branch but then continue decoding with stale cache data. This caused token divergence, especially in the FP16 configuration where a small cache difference can more easily change the next selected token.
• Fix: We changed PaKVReorder to produce two real outputs:
  output 0: updated Key cache
  output 1: updated Value cache
  The CPU plugin marks each output as in-place with its corresponding input:
  Key cache input <-> Key cache output
  Value cache input <-> Value cache output
  Therefore, no additional full-cache copy is needed.
  We also changed the EAGLE3 helper graph and fusion pass so that Key and Value cache updates are separate model outputs. GenAI explicitly binds these outputs to the external KV cache tensors.
  Now the runtime knows that the reordered K/V caches are real results that must be preserved for the next decoding step. After the fix, FP16 EAGLE3 produces the same token IDs as target-only decoding in the tested deterministic cases, and the full WWB similarity improved from 0.88574980 to 0.97614926.

Tickets:

• CVS-191948

AI Assistance:

• AI assistance used: no / yes
• If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).