Cache optimizations

Remove cache invalidation calls from the inference_process library as
they are not needed.

Fix the clean function to actually clean/flush the tensor arena after
copying the IFM data and nothing else, to ensure cache coherency before
invoking an inference on the NPU.

Change-Id: Ida7f4474c86ca3a24c15e1cb8abe84440b60fdc3
Signed-off-by: Jonny Svärd <jonny.svaerd@arm.com>