[Pass] Implement Tile Loop Fusion

[JiaqiGuoSunlune]

1. Model Loop Fusion as a Graph Partition problem with respect to a deterministic cost function (non-negatifve)
2. Solve the problem with dynamic programming
3. Enable Google Test in Tilelang
4. End-to-end testing in Python
5. Fix a bug in tile resize inference of reduction

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[firefrogliu666]

I went through this PR carefully and at this point I understand the main idea and implementation structure well enough.

My understanding is that this pass discovers adjacent lowered tile.scope_entry regions, builds a window-local dependence/planning problem, and rewrites them into shared execution-shell trees . The implementation follows a discovery -> problem -> plan -> rewrite decomposition. The core method to find the best rewriting plan is a DP algorithm that priorizes lowering write_cut_cost, shared_read_cost, live_range_penalty, and reorder_penalty.

The important part for me is that the pass reasons in terms of logical execution prefixes, normalized use_in / def_out, dependence depth rho, and resident values, instead of trying to do a purely syntactic rewrite.

I also checked the rewrite side and the test intent. Preserving local wrappers like LetStmt / AttrStmt, supporting partial-prefix fusion, and keeping planning window boundaries conservative all make sense to me. Overall, the decomposition into discovery, planning, and rewrite is clear, and the implementation direction looks sound.

Approved.