Enable model caching in intel-openvino delegate

Leverage serialization API to create IR xml and bin. Additionally,
leverage ov::cache_dir(cache_dir) to cache NPU compiled blob.

This brings down the frame dropped in camera for VC to zero.
Memory consumption also comes down. And model execution can go
through in the absence of libvpux_driver_compiler.so

BUG=b:344503269
TEST=cros_camera_service, arc-camera work as expected.
     benchmark_model, inference_diff_eval Pass

Change-Id: I63107c5d75855863b4d57c7dbe7b35ed8306d75c
Signed-off-by: Anisha Kulkarni <anisha.dattatraya.kulkarni@intel.com>
Reviewed-on: https://chromium-review.googlesource.com/c/chromiumos/platform/tflite/+/5736330
Commit-Queue: Shik Chen <shik@chromium.org>
Reviewed-by: Shik Chen <shik@chromium.org>