[OpenVINO] Fix compiled-model cache invalidation and expand test whitelist - #3029

Description of the change

ov_infer reused a compiled model whenever the device matched, so a second generate() with different input shapes ran against a graph traced for the first one. compile_model also relaxed every input dimension to -1 after tracing, advertising shapes the graph was never built for. The cache is now keyed on the input signature and static args, and the relaxation is gone so the compiled model matches what was traced.

The BLIP2 early-stopping mocks read token_ids.shape[0] as a Python int, which is None under a dynamic batch. They now use the ops.ones_like + ops.slice_update idiom the other causal-LM tests already use.

Depends on keras-team/keras#23564, which is now in keras-nightly.

Closes: #3045

• I have added all the necessary unit tests for my change.
• I have verified that my change does not break existing code and works with all backends (TensorFlow, JAX, and PyTorch).
• My PR is based on the latest changes of the main branch (if unsure, rebase the code).
• I have followed the Keras Hub Model contribution guidelines in making these changes.
• I have followed the Keras Hub API design guidelines in making these changes.
• I have signed the Contributor License Agreement.