Stateless broadcast optimization

[cavusmustafa]

• Stateless translation updated so gpu plugin can capture ov::op::internal::RoPE
• Temporary performance solution for stateless gpu: Bypass v13 SDPA entirely. Express attention manually so MatMul's NUMPY broadcast handles the GQA expansion at kernel level. K and V stay at n_heads_kv shape; the GEMM kernel reads them once and broadcasts via stride trick. We can revert this once we can utilize internal SDPA kernel which supports GQA broadcasting.

[wine99]

I believe qkv arrive as [B, n_heads, S, head_size] where B is the extra input n_seq_active, so this code does not work correctly with llama-perplexity or llama-server -np > 1.

If the ov pattern supports multiple sequences, i.e. B != 1, we can change the shape to {0, num_heads_kv, 1, -1, head_size} and set special_zero = true in Reshape. If the ov pattern does not support multiple sequences, we can set use_manual_gqa_attention to false if n_seq > 1 or manually run perplexity with GGML_OPENVINO_MANUAL_GQA_ATTN=0. Otherwise LGTM

[wine99]

FYI to run llama-server -np > 1 or llama-perplexity you need to include the commit from #199