Fix invalid weight mem reuse to resolve OOR (#32530)
+ Resolve memory re-use of weight custom format changed from abcd to
AcdB16a2b
### Description of the issue
- phi-4-multimodal-instruct wwb test returned OOR error.
- Batch size of Issued convolution layer is changed from 25 to 10 as
shape changed.
- query format-tag of the onednn convolution is changed from abcd to
acdB16a2b.
- second execution with acdB16a2b referred to weight cache of the first
execution
#### The code and line that caused this issue
- Modified re-use condition of _reorderded_weights_cache in
update_weights
#### Reproduction step and snapshot
- Reproduced by wwb
`wwb --target-model
C:\dev\byungilm\work\download\WW41_llm-optimum_2025.4.0-20128\phi-4-multimodal-instruct\pytorch\OV_FP16-4BIT_DEFAULT
--gt-data phi-4-multimodal-instruct__NAT\reference.csv --model-type
visual-text --genai --device GPU --output output_phi-4-multimodal
--verbose`
- Reproduced by benchmark
`.\benchmark_app.exe -m
C:\dev\byungilm\work\download\WW41_llm-optimum_2025.4.0-20128\phi-4-multimodal-instruct\pytorch\OV_FP16-4BIT_DEFAULT\openvino_vision_embeddings_model.xml
-d GPU -nireq 1 -data_shape
pixel_values[25,3,448,448],[10,3,448,448],patch_position_ids[25,1024],[10,1024],patch_attention_mask[25,32,32],[10,32,32]
-niter 1`
`
#### Checklist
- [x] Is it a proper fix?
- [X] Did you include test case for this fix, if necessary?
- [x] Did you review existing test that can be extended to cover this
scenario? Passed llm_bench
### Tickets:
- CVS-172761
---------
Signed-off-by: Min,Byung-il <byungil.min@intel.com>
Signed-off-by: Min, Byungil <byungil.min@intel.com>
