[testing][py_converter] Enhance py_converter to better support entire modules

[slyubomirsky]

This PR makes a few improvements to py_converter to make it more useful for fuzz testing, especially when running larger modules.

In particular, these changes are to support returning the definition of a global var directly (e.g., if you do run_as_python(main_var, mod=mod), the result will be a function corresponding to mod["main"]) and to correct two bugs in the previous implementation:

• Previously, it was not possible to insert a function into a runtime container object like an ADT. This was because the converter was simply compiling Relay functions into Python functions. This change solves this problem by registering the functions into PackedFuncs. However, another fix was also needed: Even though PackedFunc is an ObjectRef in C++, the Python bindings do not recognize PackedFuncs as Objects, so the code now calls the FFI API tuple constructor directly.
• The implementation relied on IRModule.from_expr to wrap passed in expressions in a module. However, from_expr will not overwrite the main function if one is passed in via the functions argument. Thus, if the user passed in a module that already had a main function defined, the wrapping would be done incorrectly and result in the main being copied many times. This PR corrects this error by not assuming that the name main will be available and instead constructing a new module with a reserved name for the target.

None of the cases above had been tested before (there are now tests included)

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• No users to tag found in teams: testing, py_converter _{See #10317 for details}

_{Generated by tvm-bot}

[slyubomirsky]

Please review @junrushao (or tag other reviewers)

[slyubomirsky]

Hm, this test worked in my fork. Wonder why it fails in mainline

[slyubomirsky]

The problem seems to be that this cast on mainline will not convince Python that the result is an opaque Object rather than a PackedFunc. Not sure how to change it:

TVM_REGISTER_GLOBAL("runtime.CastPackedFuncToObject").set_body_typed([](const PackedFunc& pf) {
  return Downcast<ObjectRef>(pf);
});

Any advice?

[slyubomirsky]

@tqchen helpfully advised that the issue stems from passing a list containing a PackedFunc to the constructor. Calling the ADT constructor via the FFI API directly (while a little strange) gets around this issue. That seems like a good fix that does not require disrupting other sections of the codebase, so I will implement it

[slyubomirsky]

The CI error appears to be spurious.

[slyubomirsky]

@tvm-bot rerun

[slyubomirsky]

I think there's been another spurious failure.

[slyubomirsky]

@tvm-bot rerun

[junrushao]

CC @Hzfengsy would you like to review this PR? Thanks!

[Hzfengsy]

LGTM. Thanks a lot, @slyubomirsky!