[Unity][MSC][pre M1.2] Reconstruct codegen

[Archermmt]

This is a pull request for MSC(Multi-System Compile)
RFC: https://discuss.tvm.apache.org/t/rfc-unity-msc-introduction-to-multi-system-compiler/15251/5
Tracking issue: #15233

This is the pre-2nd part of Milestone 1: Finish RuntimeManager for frameworks

To limit each PR in reviewable size, the Milestone 1 will be split into some steps:
M1.1 Add translate && codegen for torch
pre M1.2 reconstruct codegen
M1.2 Add translate && codegen for tensorflow
M1.3 Add codegen for tensorrt
M1.4 Add RuntimeManager and test with relax
M1.5 Add RuntimeManager and test with torch
M1.6 Add RuntimeManager and test with tensorflow
M1.7 Add RuntimeManager and test with tensorrt

After I test some cases in tensorrt and tensorflow, I found out that the codegen module need to be reconstructed. To make the development in tensorflow and tensorrt better, This PR change the design of codegen.

@Hzfengsy
@junrushao

[Archermmt]

@Lunderberg could you please check why the error occurs in CI test:

tests/python/relax/test_analysis_suggest_layout_transforms.py::test_non_affine_access [01:22:12] /workspace/src/relax/analysis/layout_transformation.cc:443: Warning: [LayoutInference] Failed to analyze indices [v_ax0 * v_ax1, v_ax2], error: ["Product of two iterators cannot be represented as an IterMap, occurs in v_ax0 * v_ax1"]

I found out that you meat the same error in #15810 and #15807, but no errors in #15815. Did you fixed the bug?

[Lunderberg]

I had a couple of MSC-related CI failures, but they were mainly related to expected IRModule definitions needing to be updated for changed behavior. If I run the tests/python/relax/test_analysis_suggest_layout_transforms.py::test_non_affine_access on #15815, I get the same warning, but the test still passes.

Checking against the CI output for #15815 (link), I see the same non-affine warning occurring there as well, which makes sense for a test validating the behavior under a non-affine transform.

I'm wondering if we want to downgrade the LOG(WARNING) to a DLOG(INFO), since it is the expected behavior when analyzing a PrimFunc that is not an affine transformation.

[Lunderberg]

(Made #15820, which updates the warning to an info.)

[Archermmt]

Thanks! Now I understand....