[CPU] Allow Gather on KV cache paths in MatchSdpaKvCache

[CuriousPanCake]

Details:

MatchSdpaKvCache is not executed due to the presence of Gather among the Concat's children on KV cache paths after StatefulSDPAFusion. This prevents MemoryInputSDPA creation, leaving KV states unassigned and causing
a null state crash at inference.

          ┌─────────┐               
    ┌─────┤ReadValue├────┐          
    │     └────┬────┘    │          
    │          │         │          
    ▼          ▼         ▼          
┌────────┐ ┌───────┐ ┌──────┐       
│SDPAWith│ │ShapeOf│ │Gather│       
│KVCache │ └───────┘ └──────┘       
└────────┘            //Not Expected

Tickets:

• CVS-183493

AI Assistance:

• AI assistance used: yes

Signed-off-by: Andrii Staikov andrii.staikov@intel.com

[Copilot]

Pull request overview

Fixes a CPU plugin graph-optimizer pattern match so MatchSdpaKvCache still runs when a Gather node appears as an additional consumer on KV-cache paths after StatefulSDPAFusion, preventing missing KV state assignment and a null-state crash during inference.

Changes:

• Extends MatchSdpaKvCache’s eligibility check to allow Type::Gather among MemoryInput’s children (in addition to ScaledDotProductAttention and ShapeOf).

[Copilot]

[HIGH] This change fixes a crash-prone behavioral bug in KV-cache state assignment, but the PR doesn’t add a regression test covering the newly accepted pattern (MemoryInput having an additional Gather consumer alongside SDPA/ShapeOf). Please add a CPU plugin unit/functional test that builds a model where ReadValue/MemoryInput feeds the fused SDPA-with-KV-cache path and also has a sibling Gather consumer, and asserts compilation + inference works (no null-state crash) and the KV states are assigned/usable.

[CuriousPanCake]

This PR is a WA. The rootcause was additionally investigated and a solution merged within bcbe69a