[CODEGEN][CONTRIB] Various update for CoreML codegen

[kazum]

• Update relay.ext.coremlcompiler based on the change in [RUNTIME] Introduce MetadataModule to separate code compilation/interpretation and weight initialization #5770.
• Support int32 for Core ML input and output
• Handle "run" in CoreMLRuntime::GetFunction to measure time with Module.time_evaluator.
• Return PackedFunc() when CoreMLRuntime::GetFunction handles nothing.
• Add test for each operator supported by CoreML codegen.
• Support relay.expand_dims and relay.nn.relu.

[zhiics]

aah, I haven't noticed that CoreML codegen was also using BYOC infra. Thanks for updating it. I have a question which is not relevant to this change. I see SetInput needs to allocate memory and copy from DLTensor to the local memory every time. Is it possible to pre-allocate memory or directly do zero-copy by sharing the passed tensor?

[kazum]

@zhiics Relevant changes to #5770 are:
https://github.com/apache/incubator-tvm/pull/5934/files#diff-639424334c798703f0e62ec8a5eaf779R30
https://github.com/apache/incubator-tvm/pull/5934/files#diff-271e7167e72d0f1de3aee097fa5cb5d2R232-R245

Is it possible to pre-allocate memory or directly do zero-copy by sharing the passed tensor?

We can at least pre-allocate memory. I think of doing more optimization for input tensor management in another PR :)

[zhiics]

@kazum no problem. We shouldn't do that in this PR. I was just curious if you have a plan to do that.

[zhiics]

Thanks @kazum