unify api calling

[zhaixuejun1993]

1. Create new interface "create_ov_output_tensor()" to create the output tensors
2. Fixed the issue caused by the shape change in llama-bench runtime
3. Unified the calling for this to avoid many sub code snippet in many place
4. this change is also for the future changes (hash index replace the name index)

[wine99]

3259921#diff-f03e5dcc8ad94468267d70b7a2ff7926977ecbc1395285e416ee5ad3b13fab78R357-R359

LGTM. The modified lines were originally introduced in the commit above by @cavusmustafa — could you take a look?

[cavusmustafa]

The earlier change was to fix llama-bench -d 1024 failure on NPU. After this change it fails again on my side.

[zhaixuejun1993]

The earlier change was to fix llama-bench -d 1024 failure on NPU. After this change it fails again on my side.

@cavusmustafa
the failed caused by the output shape of "result_norm" changed from 1 iter to 2 iter on prefill phase ([1, 1, 1, 2048] -> [1, 1, 0, 2048]). Will enable your solution get shape from inferrequest to avoid this kind of issue.

[zhaixuejun1993]

Verified with llama-simple & llama-bench (-d 1024). PASS.
pls review again