[Unity][Dlight] Enabling Fast and Efficient Kernel Generation by leveraging Hardware information

[LeiWang1999]

This pull request serves as an enhancement to Dlight. This update primarily focuses on incorporating hardware information to recommend tile candidates, which enables fast tuning.

Below is a brief summary of the major changes:

• Introduce dl.ApplyFastTuning Pass

• new flag for skip_simplify schedule::reindex : This addition can help avoid over-optimization of eliminating unit loops. This flag functions similarly to preserve_unit_loop.

• improve compute_inline.cc to enhance the simplification on some complex inline case (e.g. layout transform)

• simple bug fixes:

  • [TIR] Fix of inter thread reduction with shared memory prefetch #16406
  • [Unity] Improve buffer allocation for handling duplicated buffer names. #16437

Related discussion: https://discuss.tvm.apache.org/t/dlight-enabling-fast-and-efficient-kernel-generation-by-leveraging-hardware-information/16273

TODO Items of this pull request:

• provide related testing.
• code style may need guidance (e.g. remove or replace the log print; refactor some python components to cpp)
• leverage structural equal cache to avoid duplicated tuning.
• implement mma schedule template with swizzling.
• support dynamic symbolic tuning
• bring it to mlc-llm (maybe should improve our design to support dynamic symbolic).

[LeiWang1999]

we only had shared.dyn because the merge_shared_memory can only merge dynamic, currently we can also enable static shared memory merge.