[Relay][Pass] Add inline pass

[zhiics]

This PR adds inlining pass to Relay. It can help external codegen where we may want to outline the sub-functions (subgraphs) to the module level to prevent the from being optimized by relay passes. These functions can then be inlined and offloaded to external codegen tools as discussed in #4864

This transformation is also important to the Relay VM as inlining may reduce the calling overhead. We currently only greedily inline the functions that are marked inline whenever it is abled to be inlined. There is no profitability analysis available for now.

This PR is dependent on the call graph PR #4922

[zhiics]

cc @tqchen @jroesch @icemelon9 @yzhliu @anijain2305 @mbaret @masahi @soiferj

[masahi]

@zhiics can you rebase?

[zhiics]

@masahi Done. Thanks.

[masahi]

@zhiics not sure if this is one of an intended use cases, but I think adding a test case for inlining a composite function would be a useful addition. (This is not a request but soft suggestion, feel free to ignore).

[zhiics]

@masahi We have annotated functions with external compilers, no difference with composite if we need to add it. I don’t think we need to add it in this pr

[masahi]

Does this pass have any relation with InlinePrimitives pass? Do they have different use cases?
https://github.com/apache/incubator-tvm/blob/6798ba80d288e7c6132b30606b3cb70812579fe8/src/relay/backend/vm/inline_primitives.cc#L149-L160

[zhiics]

@masahi inlineprimitives works on each function scope to inline the function lifted by anf. This pass handles functions at the global scope.

[masahi]

This is a cool use of the new call graph infra, LGTM.

[anijain2305]

Thanks @masahi @zhiics This is merged