[iOS] Add optional Selective Registration of Ops (#14421)

The current iOS library is huge. Add the ability to selectively
register for the ops the tensorflow library will support. This
greatly reduces resultant binary based on the network.

A "full" arm64 build was 122MB on my machine vs one selectively
registered for SSD Mobilenet was only 93MB.

Also fixes a minor bug where the selected arch wasn't being passed
to the compile_ios_protobuf.sh script.

TEST:build_all_ios.sh -a arm64 # generates a fat binary for arm64
     build_all_ios_sh -a arm64 -g ~/Downloads/op_inference_graph.pb
        #generates a binary that is much smaller