[GPU] Fix per_layer_inputs shape in SDPAToPagedAttention for Gemma-4-e2b

[hyunback]

Problem

Gemma-4-e2b-it crashes during GPU compile_model with PagedAttention enabled.

The SDPAToPagedAttention pass merges batch and seq dims of inputs_embeds into 2D. It inserts an Unsqueeze to produce [total_tokens, 1, hidden] for downstream consumers. However, the per_layer_inputs parameter (PLE) retains its original shape [?, ?, ?, 256].

When added to the PLE projection output [?, 1, 35, 256], broadcasting on dim1 (1 vs ?) makes all downstream shapes dynamic. This propagates through every layer's residual Add. All PA query shapes become [?, ?] instead of [?, 2048].

On GPU, this causes a crash — heads_num must be computed at compile time from a static query dimension. On CPU, it silently degrades into unnecessary dynamic shape inference across 34 layers.

Solution

Set per_layer_inputs dim1=1 in SDPAToPagedAttention. This matches the PA mode convention where batch×seq are merged and Unsqueeze inserts seq=1. It is consistent with what the GenAI CB runtime already provides: [total_num_tokens, 1, num_hidden_layers, hidden_size].

All 35 PA query shapes become static. GPU/CPU compile succeeds. The broadcast-induced dynamic propagation is eliminated.

Tickets:

• CVS-190998

AI Assistance:

• AI assistance used: yes
• If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).